Comparative Methodologic and Practical Considerations for Life Expectancy as a Public Health Mortality Measure

Abstract

Introduction

Life expectancy is a public health metric used to assess mortality. We describe life expectancy calculations for US counties and present methodologic considerations compared with years of potential life lost before age 75 (YPLL-75) and premature age-adjusted mortality (PAAM), 2 commonly used length-of-life metrics.

Methods

We used death data from the National Center for Health Statistics for 2015-2017 and other health measures from the 2019 County Health Rankings & Roadmaps. We calculated life expectancy from birth at the county level using an abridged life table and the Chiang method of variance. Studentized residuals identified counties with discordant life expectancy and YPLL-75 or PAAM values. Correlations tested associations of life expectancy with key health measures (eg, smoking, child poverty, uninsured).

Results

Among 3073 US counties, life expectancy ranged from 62.4 to 98.0 years, with a mean of 77.4 years. Life expectancy was strongly and negatively correlated with YPLL-75 (r = −0.91) and PAAM (r = −0.95) at the county level. Life expectancy was also associated with other key health metrics, such as smoking, employment, and education rates, where an improvement in the health factor indicated improvement in the respective length-of-life measure. Counties with discordant life expectancy and YPLL-75 or PAAM values had differing age structures.

Practice Implications

Commonly used length-of-life metrics in population health settings are differentiated by methodological matters, such as computation complexity, data availability, and differential risk among age groups, especially among the very old or very young. The choice of metric should consider these factors, in addition to practical concerns, such as the communication needs of the audience.

Length of life, an important indicator of the health of a community, can be measured with various metrics. Premature age-adjusted mortality (PAAM) rates (deaths per 100 000 population), cause-specific mortality, and years of potential life lost before age 75 (YPLL-75) are commonly reported metrics in public health. ^1,2 Although used frequently by demographers for years, life expectancy has gained prominence in public health during the past few decades to assess and communicate length of life and current community health. ^{3 -5}

Life expectancy is the average number of years from birth a person can expect to live, according to the current mortality experience of the population. Paradoxically, the calculations and assumptions underlying life expectancy are more complex than other commonly used length-of-life measures, yet the concept of life expectancy is more easily understood than PAAM or YPLL-75 by lay audiences. ^6,7 We describe the calculation of life expectancy for US counties and summarize key methodologic considerations in comparison with YPLL-75 and PAAM to inform public health practice and research.

Materials and Methods

Data Sources

Mortality data at the county level came from the mortality files and population data came from bridged-race (July 1) population estimates, both from the National Center for Health Statistics for 2015-2017. ⁸ We derived other county-level metrics from the 2019 County Health Rankings & Roadmaps (CHRR) data set (https://www.countyhealthrankings.org), consisting of more than 50 measures of health from numerous data sources (a list of measures and data sources is available as online supplemental material). The University of Wisconsin Institutional Review Board considered this study exempt from oversight because data were at the county level and no individual private information was used.

Life Expectancy

We calculated life expectancy by applying Chiang’s methodology for an abridged life table. By extrapolating a mortality experience onto a hypothetical population, this method facilitates the comparison of the cumulative years a person can expect to live across counties with different age structures and produces reliable estimates for small populations. ^{9 -12} More specifically, Chiang’s methodology for an abridged life table first uses death and population information to calculate the risk of dying in predefined age groups. Next, these death rates are applied to a hypothetical starting population of 100 000 to determine the hypothetical number of deaths that would occur in each age group and the hypothetical number of people alive at the beginning of each age group. As the age group advances from one to the next, the population (which started at 100 000 at birth) decreases to reflect the number of deaths that have occurred during the interval. The population alive at the start of each age group and the population dying in each age group are used to estimate the total number of years the population lived in each age group. Then, these values are summed across the age groups to give the total number of years the population lived beyond each age group. Finally, the total number of years the population lived beyond the age group is divided by the population alive at the beginning of the age group, resulting in the observed expectation of life at that age. Detailed information on the variables and formulas used to calculate life expectancy and an example life table calculation are available as online supplemental material.

We calculated variance in life expectancy via the Chiang methodology with no zero-death count substitution and adjusted Chiang variance for the final age interval.^6,11-13 We considered substitution methods to compensate for the underestimation of the true variation due to zero death counts, including substitute values of 0.693 and 3.0, the Poisson means where the probability of observing zero deaths is 50% and 5%, respectively. However, this approach did not improve estimates, consistent with previous findings, ^6,12 and, therefore, was not used. Alternatively, we added a term to the variance of the final age interval based on a Poisson distribution of the death counts. This method mitigated underestimation of the overall variance from the final age group. ¹³

We performed life expectancy calculations at the US county, state, and national levels, with deaths counted in the county of residence. We reported life expectancy from birth (age 0) calculated using age intervals of 5 years, with the exception of infants and people aged 1-4 and ≥85 years (ie, <1, 1-4, 5-9, 10-14, 15-19, . . . , 80-84, ≥85). We used 5-year age groups, with ≥85 as the final age group, because this combination results in the least overestimation of life expectancy for small populations. ¹² Counties were assigned missing life expectancy values if the estimate’s SE was >25% of the estimate itself. This is known as a relative SE >0.25, a commonly used threshold for unacceptable variability. Estimates of life expectancy were also assigned missing values if the county had fewer than 5000 population years at risk in the time frame, the point at which estimates are thought to yield biased values and large errors. ^7,12

Analyses

To test the criterion validity of life expectancy, we examined associations with YPLL-75 and PAAM using studentized residuals, a statistical method to identify outliers. We identified outlier counties by discordance of >3 SDs from the mean of the regression between YPLL-75 and life expectancy or PAAM and life expectancy. We intuited that the discordance would arise from different death rates in either the very young or very old age groups, given the construction of the 3 measures. Because access to underlying death data is not available to all public health practitioners, we explored differences in community characteristics by using available demographic variables that could be considered proxy measures for these potential differential death risks. We performed t tests on county-level demographic variables of age, race/ethnicity, and population size. ^14,15 We set significance at P < .05.

We examined construct validity of life expectancy by using Pearson correlations with other measures from CHRR. Measures included self-rated health, physically unhealthy days, mentally unhealthy days, obesity, smoking, uninsured, high school graduation, children in poverty, and severe housing problems. We compared correlations with YPLL-75 and PAAM. We conducted data management and statistical analyses using SAS version 9.4 (SAS Institute, Inc).

Results

Descriptive Statistics

We calculated life expectancy for 3073 of 3142 US counties. Life expectancy ranged from 62.4 to 98.0 years, with a mean (SD) of 77.4 (3.0) years. Sixty-nine counties were assigned a missing value because they had populations of <5000 population years at risk during 2015-2017. No counties had a relative SE >0.25. The average SE for county-level estimates was 0.69 years, ranging from 0.03 to 6.60 years (data available as online supplemental material). State-level life expectancy varied from 74.8 years in Mississippi to 82.2 years in Hawaii (Figure 1). Life expectancy for the United States was, on average, 79.1 years.

Figure 1.

County-level distribution of life expectancy (A), YPLL-75 (B), and PAAM (C), US counties (n = 3142), 2015-2017. Data source: National Center for Health Statistics. ⁸ Abbreviations: PAAM, premature age-adjusted mortality; YPLL-75, years of potential life lost before age 75. Bin widths were determined by using half SD breaks, and the mean was used as the midpoint of the 12 categories for life expectancy and YPLL. Because PAAM is heavily right-skewed, the mean + 0.5 SD was used as the center point for the 12 categories.

Relationship to Other Length-of-Life Measures

Life expectancy was strongly, negatively correlated with YPLL-75 (r = −0.91) and PAAM (r = −0.95) at the county level (Figure 2).

Figure 2.

Comparison of values of life expectancy with YPLL-75 and PAAM, US counties (n = 3113), 2015-2017. Data source: National Center for Health Statistics. ⁸ Abbreviations: PAAM, premature age-adjusted mortality; YPLL-75, years of potential life lost before age 75.

Studentized residuals identified 52 outlier counties in the association of life expectancy and YPLL-75 and 35 outlier counties in the association of life expectancy and PAAM. Because the counties identified as outliers in both comparisons were similar (ie, only 5 outliers in the PAAM/life expectancy comparison were absent in the YPLL-75/life expectancy comparison), we describe only the demographic characteristics of the outliers in the association between life expectancy and YPLL-75. County demographic variables showing a significant difference between outliers and nonoutliers were the variables of percentage of the population aged <18 years, aged >65 years, or Hispanic, and total population. Compared with non-outlier counties, outlier counties with a life expectancy of <75 years (Figure 2) were significantly more likely to have a higher percentage of people aged <18 years (32.5% vs 22.2%), a lower percentage of people aged >65 years (12.2% vs 18.7%), and a smaller population (11 812 vs 107 000) (all P < .001). We found no significant difference in the percentage of the population that was Hispanic. Outlier counties with a life expectancy of ≥75 years were significantly more likely than non-outlier counties to have a higher percentage of population that was Hispanic (19.2% vs 9.3%; P = .02), a lower percentage of people aged <18 years (18.8% vs 22.3%; P < .001), a higher percentage of adults aged >65 years (23.4% vs 18.6%; P = .003), and a smaller population (21 501 vs 107 275; P < .001).

Relationship With Other Health Measures

County-level correlations showed strong, positive relationships between life expectancy and key measures of health. All associations were significant (P < .001) in the expected direction, but the strength of associations varied (Table). The magnitude of associations with life expectancy was strongest for measures of smoking, children in poverty, self-rated health, and mentally and physically unhealthy days. For life expectancy, the correlation was highest for smoking (−0.70), whereas YPLL-75 and PAAM were most highly associated with children in poverty (0.69 and 0.73, respectively).

Table.

Associations of mortality measures with key county-level health measures in a study of life expectancy, US counties, 2015-2017 ^a

Health factors	Life expectancy b	YPLL-75 per 100 000 population (counties) b	PAAM per 100 000 population (counties) b
Poor or fair health	−0.64 (3073)	0.64 (3081)	0.70 (3081)
Poor physical health days	−0.67 (3073)	0.65 (3081)	0.71 (3081)
Poor mental health days	−0.64 (3073)	0.59 (3081)	0.65 (3081)
Low birthweight	−0.48 (3029)	0.50 (3028)	0.53 (3028)
Smoking	−0.70 (3073)	0.66 (3081)	0.71 (3081)
Obesity	−0.54 (3073)	0.45 (3081)	0.52 (3081)
Uninsured	−0.23 (3073)	0.30 (3081)	0.31 (3081)
High school graduation	0.08 (2999)	−0.13 (3004)	−0.10 (3004)
Unemployment	−0.41 (3073)	0.44 (3081)	0.44 (3081)
Children in poverty	−0.65 (3073)	0.69 (3081)	0.73 (3081)
Severe housing problems	−0.07 (3073)	0.15 (3081)	0.10 (3081)

Abbreviations: PAAM, premature age-adjusted mortality; YPLL-75, years of potential life lost before age 75.

^aAll correlations are significant at P < .001 using Pearson correlations. All values are r (number of counties). Numbers in parentheses correspond to the number of counties used in the analysis. Data sources: Health factor data are from County Health Rankings & Roadmaps, ⁵ and mortality data are from the National Center for Health Statistics. ⁸

^bLife expectancy is the average number of years a person can expect to live from birth using life tables. YPLL-75 is the number of years of potential life lost before age 75 per 100 000 population (age adjusted). PAAM is the number of deaths among residents aged <75 per 100 000 population (age adjusted).

Discussion

Length-of-life measures are essential indicators of population health. We described the calculation of life expectancy, the most methodologically complex length-of-life measure, and comparative analyses uncovered several important methodological nuances that are relevant in choosing between measures. We discuss several potential reasons for the discordance between measures and measure limitations, and we raise important communication considerations and implications for public health practice.

Discordance Between YPLL-75 or PAAM and Life Expectancy

We found a strong, but not perfect, correlation between life expectancy and YPLL-75/PAAM. We investigated potential reasons for this discordance by examining the associations between life expectancy, YPLL-75, and PAAM and important demographic variables, which served as proxies for differential risks of death among counties. The primary explanation for the discordance was the influence of differential risk of deaths among age groups between measures.

Life expectancy and measures of premature death differ in their inclusion of adults aged ≥75. By including deaths from all age groups, discordance between life expectancy and YPLL-75/PAAM values can occur if the risk of death in a population at or after age 75 is substantially different from the risk of death before age 75. This discordance can occur, for example, in counties with large, affluent retirement populations, where the oldest age groups may have a differential risk of death that is lower than the rest of the population because of the migration of healthier, wealthier, older adults into these communities. Also, differential risk in the oldest age groups may occur in counties with larger Hispanic populations from issues related to the “Hispanic paradox,” where mortality rates among the Hispanic population are artificially deflated as a result of the oldest and sickest members returning to their country of origin to die. ^15,16 Instances of differential risk can be seen in the outlier counties with a life expectancy of ≥75 years (Figure 2), where life expectancy values are more dispersed than YPLL-75 values. Because differential risk of death matters, it is unsurprising that outlier counties have smaller populations, on average, than non-outlier counties. In these counties with smaller sample sizes, a small change in deaths due to chance can contribute to discordance.

On the other hand, by weighting deaths in younger age groups more than older age groups, counties’ premature death values are strongly influenced by infant and child mortality rates. This unequal weighting will lead to discordance in life expectancy and YPLL-75/PAAM in counties with an especially high risk of death in the youngest age groups. Some examples of this discordance can be seen in outlier counties with a life expectancy of <75 years (Figure 2), where YPLL-75 values are more dispersed than life expectancy values. Accordingly, these outlier counties have higher rates of infant mortality and death among age groups aged <75 than non-outlier counties.

Two additional sources of discordance are the inclusion of different age groupings and suppression criteria between measures. Five-year age groups were used for life expectancy, whereas 10-year age groups were used for YPLL-75 and PAAM, which can lead to discordant results if differential risks of deaths or age structures occur between these groupings. Although 5-year age groups are often recommended for calculating life expectancy in public health, it is possible to use 10-year age groups for exact comparisons or to account for small numbers. ¹² Our analysis used data from CHRR, a common data set used by public health practitioners; as such, we did not perform further analysis of various age groupings. Differential suppression criteria, based on death counts for PAAM/YPLL-75 and population size for life expectancy, can also systematically lead to different patterns of missingness for counties on the basis of their populations. For example, life expectancy is more likely to capture counties with larger and healthier (fewer death counts) populations, whereas YPLL-75/PAAM are more likely to capture counties with smaller and less healthy (more death counts) populations.

Relationship With Other Health Measures

The association between life expectancy and other health measures was consistent with associations reported in other studies. ^{17 -24} Associations were in the expected direction, such that an improvement in the health factor indicated improvement in the respective length-of-life measure. The magnitude of associations for health factors with life expectancy was on par with associations with YPLL-75 and PAAM, measures included in CHRR since 2010. ²⁵ However, some nuanced differences occurred. Health behavior metrics of smoking and obesity were slightly more associated with life expectancy than with YPLL-75/PAAM, whereas social and economic factors and clinical care measures were slightly more associated with YPLL-75/PAAM than with life expectancy. Some variation between these associations was expected given that life expectancy includes all deaths, whereas premature death measures only include deaths before age 75. These results seem to indicate that people who live in counties with poorer social and economic outcomes are more likely to die prematurely than people who live in counties with better social and economic outcomes. ²⁶ Alternatively, poor health behaviors may exercise more long-term effects on health, resulting in chronic health problems, such as chronic obstructive pulmonary disease or lung cancer from smoking or heart disease from obesity, diseases that generally affect older segments of the population.

Measure Limitations

The underlying data used in this analysis had several limitations. First, several limitations have been documented in the collection of vital statistics and population data from census, such as incomplete death certificates. ^{27 -30} In addition, ancillary examinations of disaggregated, county-level data for length-of-life measures, such as by race/ethnicity, where death counts of zero within age groups are common in small counties, confirmed the tendency for overestimation in small populations. ^6,12 The calculation of life expectancy is more prone to bias from death counts of zero within age intervals than YPLL-75/PAAM as a result of the entire hypothetical population moving through to the subsequent age group. In addition, life expectancy is more likely than YPLL-75 to feature zero death counts because it uses 5-year age intervals as opposed to 10-year age intervals. For small counties and any disaggregation to smaller groupings, death counts of zero within age groups are likely to lead to an underestimation of the true variation. However, as stated previously and found in other analyses, estimates of SEs were found to be robust, with neither of the alternate substitution methods producing better estimates. ^6,12 Further research on methodologic considerations with small samples, such as Bayesian or random effects modeling procedures to improve reliability and decrease bias with small estimates, sensitivity to population demographic shifts, and suppression of highly improbable values, among others, warrants further consideration. ⁹

Second, the issue of smoothing within age intervals assumes that people die, on average, in the middle of the interval. This assumption may impose slight errors in the estimates. For example, if a death falls in the 5-9 years age group, it is assumed that the death, on average, occurs at age 7. If deaths occurred more often in a county at age 5 than at age 7, an overestimation of the true mortality experience could result. For life expectancy, smoothing within age intervals is particularly concerning in the ≥85-year age group, where differences in age structure can lead to misleading results by implying similarity in death risks. ⁶  For example, consider 2 counties with the same risk of death in the ≥85-year age group, 1 county’s population concentrated at age 85 and the other’s concentrated at age 100. In reality, these counties have much different expectations of life but will have the same calculated value of life expectancy.

Practice Implications

Measuring the average mortality experience of a community can be achieved using various metrics. Although life expectancy, YPLL-75, and PAAM all measure mortality, several considerations and opportunity costs exist in the methodological computation and communication of each measure that guide the choice of one measure over another. In choosing a metric, several important considerations include the intended purpose, audience, and communication strategy, as well as the community’s values and norms for explicitly addressing disparities in methods and measurement. For example, mortality measures weight age groups differently, and a county interested in reducing death rates of the very young (eg, child mortality) might choose a different metric than a county interested in reducing all cause-specific deaths (eg, car accidents). In addition, understanding data availability, methodologic capacity, and communication needs can help guide choice in metric.

PAAM rates are the easiest to calculate mathematically, requiring only age adjustment to compare across counties. However, reporting this measure as a rate per 100 000 population can be difficult to interpret, especially for smaller communities with populations of <100 000. YPLL-75 is also age adjusted and adds in the computation element of years of potential life lost in the final rates, sometimes confusing lay audiences. As a result, additional importance is applied to deaths at young ages. For example, in YPLL-75, an infant death (74.5 years lost) counts as more than 10 times more important than a death between age 65 and 74 (5.5 years lost). The increased value of a younger life is similarly assigned in life expectancy. Life expectancy is often the easiest metric to understand, with the caveat that it measures expectancy from birth rather than from a person’s current age, a common misconception. Yet life expectancy presents many calculation challenges, because advanced statistical software is often needed to efficiently create the lifetables. To alleviate some of the difficulty and confusion in calculating these metrics, CHRR generated a tool to increase data availability and fluency by assisting in calculations of life expectancy, YPLL-75, and PAAM. ³¹

Although understanding the methodologic strengths and limitations of population health metrics is essential, it is also important to consider the interpretation of population health metrics and the messages that can be communicated in an understandable way. We briefly describe these considerations and acknowledge that more translational and applied research is needed to support public health practitioners. Mortality measures differ in their interpretation as focused on either the individual or the community. Life expectancy is often misunderstood to reflect one’s own experience (ie, how long can I expect to live?), ³² despite it actually being a community-level measure. YPLL-75 and PAAM, on the other hand, are more clearly understood as community metrics because they are reported as either combined years of life lost or total deaths per population. The inclusion of deaths in all age groups or only deaths among people aged <75 also has implications for communication. Whether one wishes to describe the mortality of the entire population (eg, life expectancy) or only those deaths that are considered premature or preventable (eg, YPLL-75 and PAAM) will guide the choice of metric. ^{32 -34} Each choice will differ based on the values and needs of the community and researchers. These considerations warrant further exploration and reflection in future research.

Conclusions

Length of life can be captured using numerous methods and measures. Each has methodologic and practical strengths and limitations that are important to consider when advancing unique research and practice priorities. Life expectancy is easily understood by users of public health data and has many methodological strengths. Other commonly used mortality metrics, YPLL-75 or PAAM, can focus attention on mortality among younger age groups in a population but may be harder to communicate. However, in certain situations, including populations with differential mortality risk among age or racial/ethnic groups in a community, measures are misaligned across these mortality metrics. These situations underscore the need to consider several analytic and applied factors in the choice of a measure and methods, such as the intended purpose, audience, and communication strategy; community values and norms for explicitly addressing disparities; and analytic capacity for measure development.