Falling behind: life expectancy in US counties from 2000 to 2007 in an international context

Modeling approach

There are three broad approaches for estimating health outcomes for small areas where directly observed data are unreliable or unstable due to small numbers of observations: pool multiple years of data, borrow strength across geospatial units, or use structured relationships with covariates (Srebotnjak T, Mokdad AH, Murray CJL: A novel framework for validating and applying standardized small area measurement strategies, submitted). Our method integrates these three approaches to measure mortality for each county and year. Specifically, we used a mixed-effects Poisson regression with time, geospatial, and covariate components. Poisson regression fits count outcome variables, e.g., death counts, and is preferable to a logistic model because the latter is biased when an outcome is rare (occurring in less than 1% of observations). The model is specified below:

where y _rjt is the death count for race r within county j in year t. Income _jt is county per-capita income for year t. Education _j is the percent of adults within county j having completed high school in the 2000 census data. Race is a dummy variable for three race groups (white, black, and other). We've grouped Asians and Native Americans into a single category to reduce the sensitivity of the model to known racial miscoding in population and death counts. is the geospatial component, calculated as the average of the posterior mode of the county random intercept for counties adjacent to county j to account for residual spatial patterns. The values for are derived from running as a prior step the same model without the geospatial component to derive the posterior values of the county random effect. μ _j is the posterior value of the county random intercept. t is the calendar year of mortality, and γ _j is a random slope on time for each county. This specification allows mortality in each county to have a unique trend. The county population size affects the contribution of the random components on death counts, leading to more emphasis on recorded death counts when predicting mortality for larger counties.

The model was estimated separately by sex and five-year age groups because the magnitude of the county random effect varies by age. The outcome of the analysis is a predicted age-, sex-, and race-specific death count for each county in the year of analysis. We used these counts, together with corresponding population figures, to calculate sex-specific and sex-/race-specific life expectancy for each county [9]. We used the method proposed by Coale and Guo to estimate the years lived in the terminal age group of the life table [10]. To produce estimates for 2000, we use data from 1995-2000 inclusive, and for 2007, we use data for 2002-2007 inclusive.

Uncertainty in county life expectancy was calculated through simulation by drawing repeatedly from the posterior distributions of the race-, age-, and county-specific death counts.

International comparisons

To show county mortality in an international context, we compared county life expectancy to an international life expectancy "frontier" time series, which we define as the average life expectancy of the 10 countries with the lowest mortality for each year from 1950 to 2010. We can use this series to calculate how many "years behind" each county is when compared to the frontier. For instance, if a county has a life expectancy in 2000 that is closest to the average life expectancy of the 10 leading countries in 1980, then we can say that this county is 20 years behind the international frontier. We calculate the number of calendar years behind each county is in 2000 as well as in 2007. This approach has been used in previous work on socioeconomic and health disparities [11–14].

Units of analysis

The 3,147 US counties were arranged into 2,357 merged county clusters, each consisting of a single county or multiple counties. There were two reasons for this merging. First, merged county units are consistent over time even when there have been changes in county definitions and borders. Second, annual death counts in many counties remain too small for stable estimates. Merging counties overcomes this limitation. Counties with fewer than 7,000 males or 7,000 females were joined with contiguous counties in the same state of similar size, income, and percent of population that is black or Native American, until the cluster met this population cutoff.

Data sources

Mortality data, including county of residence and cause of death certified and coded according to the International Classification of Diseases system, were obtained from the National Center for Health Statistics (NCHS). Standard public-use mortality files do not include geographic identifiers for deaths in counties with fewer than 100,000 people. We obtained county identifiers for all deaths for years 1959 to 2007 through a special request to the NCHS. County populations by age and race prior to 1990 were accessed through the US Census Bureau, and for 1990 and later, through the NCHS [15]. Estimates of county per-capita income were taken from the Bureau of Economic Analysis. Data on educational attainment came from the 2000 census. International estimates and projections for population and deaths were taken from the United Nations Population Division.

Model validation

All analyses were done using Stata version 10.1.

Limitations

Our findings are consistent with earlier analyses that had considered disparities across all counties, by region, and by community socioeconomic status [5, 8, 22, 26, 30–33]. Nevertheless, we found that applying the models to populations of fewer than 7,000 was associated with increased error; this, in turn, limits the granularity of the analysis that can be undertaken. There is an inherent trade-off between timeliness of estimates for small areas and the extent to which one borrows strength from structured covariates and geospatial relatedness. In settings where policy or adverse social trends may have a rapid impact on mortality, these approaches may underestimate the pace of change in mortality. Our estimates for 2007 depend on the validity of estimates of population by age, sex, and race by county, which are forecasted based on the decennial census and other data sources on births, deaths, and migration. The further away from the census year the more inaccurate these figures may become. The 2010 census round will provide a more robust assessment once it is released. Issues of age misreporting, especially at older ages, may confound the estimates of life expectancy, especially for black populations [34, 35]. There are also issues of race/ethnicity classification on the death certificate compared to self-reports in the census that could affect our estimates of white and black life expectancies at the county level [36]. It is conceivable that the observed life expectancy disparities and disparity trends could be partially explained by migration. However, the best available data on county-to-county migration rule this out as a plausible explanation [8]. The NCHS has revised US life expectancy estimates downwards [37] based on examination of Medicare data, which provides one mechanism to deal with differential age misreporting on the census and death certificates. In this study, we have not been able to incorporate these types of modifications into local estimates of life expectancy and, as such, we may be overestimating local life expectancies in the US.

Interpretation of findings

While documenting the pace of relative global decline and rising disparities is novel and may surprise some, the quantitative findings based on national data will be disputed by few. In contrast, interpreting how the US came to be in this position and what to do about it will continue to be vigorously debated. We believe it is worthwhile to examine the explanations that may help inform potential solutions, a debate that may take a long time to be resolved. While this debate continues, we believe there is sufficient evidence to identify some practical actions that can, in part, reverse this trend. The debate on the causes of poor performance will focus on three sets of factors: the social, cultural, and physical environment; modifiable behaviors, diet, and metabolic risk factors; and the performance of the health system. The roles of these factors are, of course, not mutually exclusive, as the same death can be related to social and material deprivation, risk factor exposure, and the failure of the health system.

Strong relationships have been documented between race/ethnicity, individual or community income, income inequality, and mortality in the US. While these factors convincingly affect mortality, they do not fully capture the performance variation in the US. Americans in counties with above median income ranged from being 16 years ahead of the frontier to 47 years behind. Moreover, between 2000 and 2007, 85% of these counties fell further behind the international life expectancy frontier. These findings confirm at the local level similar observations found for advantaged groups nationally [38]. Any analysis of causes of disparities will draw substantial attention to poverty, inequality, race, and ethnicity, but some of the poor performance and falling performance must be related to other factors [3].

How much of the poor performance of the US is due to differences or less favorable trends in critical risks to health such as tobacco smoking, hypertension, diabetes, physical inactivity, obesity, LDL cholesterol, diet, and alcohol? At the national level, these risk factors together lead to close to one million premature deaths [39]. If the leading four risk factors were addressed (smoking, high blood pressure, elevated blood glucose, and adiposity), life expectancy in 2005 would increase 4.9 and 4.1 years, respectively, for males and females. Disparities across eight race-county groupings would reduce by approximately 20% [39]. Given that risk factor exposures vary by county, and based on evidence from state-level analysis that risk factor exposures are larger in places with higher mortality rates [40–42], we would expect that addressing these risk factors would also tend to narrow disparities. An analysis that takes into account county exposures will be critical to fully understand the potential to reduce disparities through preventable causes of death. This, however, will require improving the measurement of exposure to leading risk factors at the local level.

How much better would the US and US counties perform if the US had had a high-performance health system? The answer rests on three dimensions that inherently underlie any analysis of health systems. First, over what duration do we assess a high-performance health system? If performance over a long duration is considered, then causes such as tobacco, road traffic injuries, and HIV might have been largely prevented or substantially mitigated. A shorter duration perspective, on the other hand, assigns a more limited scope and role to health system performance. Second, a high-performance health system could not have taken action until the scientific basis for action was established. Until tobacco was demonstrated to be a hazard, one cannot blame the health system for not taking action. Third, once the scientific basis of actions was established, to what extent should a high-performance health system have taken action? Once tobacco consumption was identified as a major risk in the 1950s and 1960s, should a high-performance health system have pursued all means to reduce consumption? Or should it have only provided information, taxed tobacco, or banned smoking in public places? Many of the debates on the extent to which the US health system is to blame for poor outcome performance turn on the scope and intensity of science-based action, which have an ideological dimension. Some in the US favor a narrow view of the duration and scope of action for a high-performance health system. These proponents emphasize treatment of disease or pharmacological management of risks. Others take a broader view [43]. Forging a consensus view on mortality attributable to a low-performance health system may be challenging.

What can be done to address the poor - and worsening - national and local performance of US communities? The US health care reform debate has focused on three strategies: extend insurance to all, improve quality of medical care for those who get sick, and focus on preventable causes of death [3]. Published studies estimate that 44,789 deaths out of 2,401,584 over age 18 in 2005 are attributable to a lack of health insurance [44, 45]. These figures may be underestimated by not taking into account the fact that insurance coverage is lowest in communities with the highest mortality rates. Even taking into account such underestimation, the number of deaths attributable to lack of insurance is dramatically too small to explain much of the poor international performance and disparities in the US.

Quality of care for disease events varies substantially across the US [46, 47]. An extensive literature highlights differences in quality as a function of race/ethnicity, income, and geography [5, 46, 47]. Improvements in quality would certainly have an impact on national life expectancy and on disparities, but there are few studies that have quantified these effects. Comparisons for specific outcomes, including breast and prostate cancer survival and acute myocardial infarction, suggest that the US, on average, has higher quality than many of the countries with better health outcomes [48, 49]. Better outcomes for cancer may be influenced by the nonrepresentativeness of the Surveillance Epidemiology and End Results (SEER) cancer registration system and the concentration of some cancers in ages over 65 with near universal health insurance through Medicare [50, 51]. For other conditions such as diabetes, however, the US has worse outcomes in some studies [52]. The Organisation for Economic Co-operation and Development (OECD) quality indicators project [53] is attempting to generate comparable measures for 541 indicators, but to date, the data have been plagued with definitional and measurement issues. To put quality-of-care issues in full light, we argue that a more comprehensive attempt to assess mortality attributable to low quality of care in the US and the impact of low quality of care on disparities should be undertaken.

Addressing leading preventable causes of death could dramatically improve the international performance of a large fraction of US counties for both males and females. What can the US health system do to realize these potential health improvements? Risks can be divided into those requiring concerted national action or community action or those that can be addressed through primary care. National, state, or even local policies [54] may be effective for banning trans-fat and regulating salt in packaged and prepared food, tobacco and alcohol taxes and control, increasing financial and physical access to healthier diets such as omega-3 fatty acids and fruits and vegetables, and authorizing the use of incentives by employers, insurers, and others for risk factor modification. Community intervention may be important for promoting physical activity and tailoring screening for hypertension, blood sugar, and cholesterol to local culture and context. Expanded and enhanced primary care can be the key locus for more aggressive management of hypertension, cholesterol, blood sugar, and personalized interventions for tobacco and alcohol. Major limitations to prioritize preventable causes of death include the need for more primary care physicians [55, 56] and implementation of research efforts to improve adherence. A health system push on preventable causes of death would not be easy, but it is a target that is technically possible and could make a major impact on US health and life expectancy rates at the national and local levels.

What could motivate people, communities, and providers to have an increased focus on preventable causes of death? Some risks can be tackled through national legislation, such as banning trans-fat in manufactured foods or increasing federal taxation on tobacco. However, we believe the combination of measurement, incentive-based financing, and local innovation will also be essential. Local measurement of the baseline level of key risks and their trends can help set priorities and evaluate performance. Given the diversity of demography, epidemiology, physical infrastructure, and health system organization at the local level, a single national solution may not be the most effective for all risks. What will work to increase the effective coverage of hypertension management in Native Americans on the Pine Ridge and Rosebud reservations and in Hispanic communities in Miami may be very different. Local innovation for addressing some preventable causes of death can be harnessed by using national and state funding to pay communities for risk reduction. The experience of the GAVI Alliance is instructive. Results-based financing is feasible, but it is imperative that measurement is undertaken independent of those with a stake in the results [57]. Given the poor performance of the US on health outcomes, a performance that is worsening each year, it is time for new thinking targeted to where the biggest impact can be made on health outcomes.