http://orcid.org/0000-0003-2423-1834
Figures
Abstract
Purpose
Despite gains in life expectancy between 1992 to 2012, large disparities in life expectancy continue to exist in the United States between subgroups of the population. This study aimed to develop detailed life tables (LT), accounting for mortality differences by race, geography, and socio-economic status (SES), to more accurately measure relative cancer survival and life expectancy patterns in the United States.
Methods
We estimated an extensive set of County SES-LT by fitting Poisson regression models to deaths and population counts for U.S. counties by age, year, gender, race, ethnicity and county-level SES index. We reported life expectancy patterns and evaluated the impact of the County SES-LT on relative survival using data from the Surveillance Epidemiology and End Results (SEER) Program cancer registries.
Results
Between 1992 and 2012, the largest increase in life expectancy was among black men (6.8 years), however there were still large geographical differences. Life expectancy was highest for Asian or Pacific Islanders (API), and lowest for American Indians and Alaskan Natives (AIAN). In 2010, life expectancies by state ranged from 73 to 82 years for white males, 78 to 86 years for white females, 66 to 75 for black males, and 75 to 81 for black females. Comparisons of relative survival using National LT and the new County SES-LT showed that relative survival using County SES-LT improved relative survival estimates for some demographic groups, particularly in low and high SES areas, among Hispanics and AIAN, and among older male cancer patients. Relative survival using County SES-LT was 7.3% and 6.7% survival points closer to cause-specific survival compared to the National LT relative survival for AIAN and Hispanic cancer patients diagnosed between ages 75 and 84 years, respectively. Importantly, the County SES-LT relative survival estimates were higher in lower SES areas and lower in higher SES areas, reducing differences in relative survival comparisons.
Conclusion
The use of these new socio-economic life tables (County SES-LT) can provide more accurate estimates of relative survival, improve comparisons of relative survival among registries, better illustrate disparities and cancer control efforts, and should be used as default for cancer relative survival using U.S. data.
Citation: Mariotto AB, Zou Z, Johnson CJ, Scoppa S, Weir HK, Huang B (2018) Geographical, racial and socio-economic variation in life expectancy in the US and their impact on cancer relative survival. PLoS ONE 13(7): e0201034. https://doi.org/10.1371/journal.pone.0201034
Editor: Iratxe Puebla, Public Library of Science, UNITED KINGDOM
Received: April 21, 2017; Accepted: July 6, 2018; Published: July 25, 2018
This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Data Availability: The authors do not own any of the data underlying this study. The data can be accessed using the following information: 1. County level mortality (deaths) data belongs to the US Centers for Disease Control and Prevention (CDC) National Center for Health Statistics (NCHS). The NCHS mortality data is a restricted use file because of confidentiality requirements. The mortality data for the period used in our study (1992-2012) is available in 2 Compressed Mortality File (CMF), (CMF 1989-98 Series 20 No. 2E and CMF 1999-2014 Series 20 No. 2T). Information on how to access the files are available at https://www.cdc.gov/nchs/data_access/cmf.htm#database. The data can be requested from NCHS (https://www.naphsis.org/research-requests). Access to and use of mortality from the National Center for Health Statistics (NCHS) requires the approval of a research review committee. The NCI Surveillance Research Program (SRP) has a Data User Agreement (DUA) with NCHS that permits SEER and SRP staff to use the compressed mortality files. We are not allowed to re-release the compressed mortality files. However, the DUA establishes that SRP can provide US Mortality data to SEER*Stat users in client-server mode via a secure password accessible server located at NCI. The data can be obtained at the county level and aggregated by 5 years age and 3-year calendar groupings and any cell with fewer than 10 deaths will have missing counts. The mortality rates linked to county level SES index data by 5 years age groups and 3-year period calendar years and dictionary are available at: https://seer.cancer.gov/expsurvival/yr1992thru2013.stateses/mortality.csv; https://seer.cancer.gov/expsurvival/yr1992thru2013.stateses/mortality.html 2. The results, i.e., the expected life tables data and dictionaries are available at the URL https://seer.cancer.gov/expsurvival/yr1992thru2013.stateses/expected.survival.csv, https://seer.cancer.gov/expsurvival/yr1992thru2013.stateses/expected.survival.html 3. Expected, relative survival and cause-specific survival data were used to validate the estimates. A signed SEER Research Data Agreement form is required to access the SEER incidence and survival data to protect identities of cancer patients (https://seer.cancer.gov/data/sample-dua.html). The data can be accessed through SEER*Stat software. The SEER*Stat sessions to obtain the data are included as Supporting Information files, to facilitate researchers to create the data once the data agreement is signed and they have access to the SEER*Stat software.
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Life tables (LT) are an important tool for calculating life expectancies [1–3] and also for the calculation of relative survival[4–6]. Relative survival is the standard method for reporting of survival from cancer registry data, as it does not rely on cause of death information from death certificates which may be missing or misclassified [7]. Relative survival is calculated as the ratio of cancer patients observed all-cause survival to the “expected” survival these patients would have in the absence of a cancer diagnosis. The expected survival, also denoted as background mortality, is calculated from population LT by matching each cancer patient in the study to their respective LT by characteristics that may affect their chances of dying from other causes.
The default LT currently used to report relative survival from the Surveillance Epidemiology and End Results (SEER) Program data [8] are national LT by sex, individual ages 0–99, race (whites, blacks and other races combined) and individual calendar years 1970–2011 [9]. These LT, herein referred to as US-LT, were constructed from decennial and annual LT from the National Center of Health Statistics (NCHS) for all races [10], whites and blacks and from NCHS mortality data for other races. Accuracy of relative survival crucially depends on how good LT represent the background mortality for the cohort of cancer patients. A recent study [11] showed that LT by state for the white and black populations captured some of the geographical variability in non-cancer cause mortality and improved relative survival calculations for younger ages but were biased for older ages. Additionally, there is evidence of variations in health status and mortality in the U.S. by geography, and by SES within the same race group[12]. Thus, national LT that do not account for variations in mortality by geography, by SES, or by race may lead to biased relative survival estimates [6, 11]. As more cancer registries in the U.S. are beginning to report relative survival it is important to have LT that represent each registry background mortality for their population to have more fair comparisons of relative survival[13]. National LT representing the average US mortality may overestimate differences in relative survival between groups of cancer patients with different mortality patterns compared to the national average. Average national LT overestimate the expected survival in more deprived areas, and underestimate expected survival in less deprived areas. Since expected survival is in the denominator for relative survival calculations, the consequence is underestimated (lower) and overestimated (higher) relative survival in more versus less deprived areas, respectively, and an increase in differences.
The goal of this study is to present a more comprehensive set of LT that more accurately represent the varying mortality patterns in different populations in the U.S. with respect to geography (county of residence), SES at the county level and individual characteristics race and ethnicity, calendar year of death, and sex. To do this, we used U.S. mortality data from 1992 through 2012 linked to SES indicators at the county level. We estimated LT by sex, calendar year (1992–2012), race (whites, blacks, Asian or Pacific Islanders (API), American Indians and Alaskan Natives (AIAN), and Hispanic-origin), state, and county-level SES as data allowed. We summarized LT by calculating life expectancies and describing patterns by state and calendar year trends by sex and race. We investigated the impact of the County SES-LT on expected and relative survival by comparing with estimates using the US-LT.
Relative survival is the preferred method to report and compare survival between different registries and countries [14] because cause of death is often unavailable, misclassified, or subject to variability on its accurate determination[7]. When cause of death is available and accurate, cause-specific survival is an alternative measure to quantify survival associated with a cancer diagnosis. An algorithm that more accurately attributes a cause of death to cancer [15] has made it possible for SEER to calculate cause-specific survival. Comparisons of relative survival and cause-specific survival are challenging because both measures are subject to bias[4].
To minimize bias, we compared survival for cancer patients diagnosed with any cancer. Because this is a broad group of cancer patients they are more likely to have similar background mortality as the general population and be less affected by bias due to cause of death determination. We hypothesized that the relative survival closer to cause-specific survival reflects the more accurate life table and relative survival.
Data and methods
County level mortality data
We used counts of deaths from the NCHS and populations from the U.S. Census Bureau available through SEER*Stat software [16] by county, single year age at death (30 to 84 years), race, sex, and calendar year 1992–2013. County is the smallest geographical area for which mortality data are available. We created mutually exclusive race and ethnicity groups, herein referred as race groups: Non-Hispanic (NH) White, NH Black, NH AIAN, NH API, and Hispanics (hereafter we exclude the NH prefix when referencing the race groups). Hispanic ethnicity includes all race categories. Because of misclassification errors of AIAN race in death certificates [17], we restricted the AIAN data to mortality rates from Contract Health Service Delivery Area (CHSDA) counties. The CHSDA counties in general contain federally recognized tribal land or are adjacent to tribal land and have health services for the AIAN populations supported by the Indian Health Service. Restricting analyses to CHSDA areas reduces AIAN misclassification on death certificates [17] and produces more accurate estimates of mortality for the AIAN populations.
County level SES index
NCHS mortality data do not contain individual level SES for deceased individuals. Therefore, we used an ecologically-defined SES index linked to mortality data at the county level. We obtained SES characteristics at the county level from the U.S. 1990 and 2000 decennial censuses and from the American Community Survey (ACS) 5-year estimates for 2005–2009, 2006–2010 and 2007–2011, used to represent the average SES in years 2007, 2008 and 2009 respectively. The SES index was previously developed and validated [18]. It used factor analyses to construct a single SES index that included poverty, unemployment, occupation, income, education, and housing characteristic [18]. We used extrapolation methods to estimate SES for the missing years, for example, the SES index for years 1995 through 1999 were estimated by extrapolating the 1990 and 2000 SES indexes. We ranked US counties from lowest to highest based on the SES index, and created equal quintiles based on combined population size: Q1 (lowest SES) were counties with 20% of the US population with the lowest SES index and Q5 (highest SES) were counties with 20% of the population with the highest SES index.
LT modeling and life expectancy calculation
We fit Poisson regression models to the log of mortality rates to estimate LT separately for men and women and each race. We used 3-year grouped mortality rates (1991–1993 for 1992, …, 2012–2014 for 2013) to provide smoothed and more stable estimates. Age and calendar year were modeled as spline functions to capture non-linear effects. Because of small numbers of deaths for younger ages and populations not being available for single ages at death for ages 85 and older, we restricted the Poisson regression model to the log of mortality rates for ages between 30 and 84 years.
Let D(i,age,year|s,r,A) and P(i,age,year|s,r,A) be the number of race r and sex s deaths and population at county i in area A by age (age = 30, …., 84) and year where year represents the midpoint of the 3 calendar years. The model assumes that the numbers of deaths follow a Poisson distribution with mean that is the product of the population and the mortality rate in a given cell, D(i,age,year|s,r,A)~Poisson[P(i,age,year|s,r,A)λ(i,age,year|s,r,A)]. The models varied by geographic area (state, region, and national) and the inclusion or not of the SES index as a covariate depending on sufficient numbers of deaths and population counts for each race. SES was included either as 5 level quintiles or 2 level grouped into low SES Q1-Q3 vs. high SES Q4-Q5. This grouping was used to maximize SES differences and ensure sufficient number of deaths and populations in the two groups. For each area (state, region or national), race and sex, the log of mortality rates was modeled as a spline function of continuous age and calendar year. The models are:
- including SES quintiles as a 5-level covariate
where county i belongs to the respective area, f and g are restricted cubic spline functions.
- including SES as a 2-level covariate
- without SES, ln[λ(i,age,year|sex,black)] = f(age)+g(year).
In all models the nodes for the spline function of age are fixed at 32, 40, 55, 70, 77, 82 and for year are 1992, 1997, 2002, 2007, and 2012. We used SAS PROC GENMOD to fit the Poisson models.
To estimate LT for ages 0 to 34 ears and 85 to 99 years we used an adjustment based on the race, sex and year-specific NCHS decennial US-LT [19]. The NCHS decennial LT have improved estimates for these age groups because they include extra data on births and better information on age at death for the very old from linkages to Medicare data [19]. The idea of the adjustment is to keep the level of the modeled LT as estimated for ages 35 and 84 and to use the form of the mortality rate by age from the US-LT to extrapolate beyond those ages. Let λUS(age,year|sex,race) be the probability of dying between age and age+1 from decennial US-LT and the state, region or national estimated LT for the respective race and sex. For ages a<35 and a>84 we adjusted the estimates as below,
and
For whites and all races combined, we fit the models for each state with county SES included as a covariate with 5 levels. The models varied for the other race groups depending on sufficient deaths and populations counts at each state. The models were: state and no SES, state and 2-level SES. For states with small populations for the respective race, LT were estimated using their respective regions and 5-level SES for Blacks and Hispanics and national with 2-level SES for API and AIAN. The models used for each state and race combination are shown in S1 Table.
To compare the County SES-LT and summarize the effects of year, sex, race, state, and county-level SES we calculated life expectancy up to age 99 using standard LT methods [20].
Comparisons of expected and relative survival
SEER collects clinical, demographic, and vital status information on all cancer cases diagnosed in defined geographic areas. Data included in this report are from SEER-18 registries (2000–2012) (November 2015 Submission) obtained using the SEER*Stat Version 8.3.2 software [21] covering approximately 30% of the US population. Relative survival is defined as the ratio of overall survival (all causes of death) by the expected survival in a comparable group of cancer free individuals and represents the excess mortality from a cancer diagnosis. Currently expected survival is estimated from US-LT matched to the group of cancer patients by age, sex, race, and calendar year. We used the Ederer II method to calculate expected survival [22, 23]. The new County SES-LT were incorporated into SEER*Stat software. Relative survival calculations match individuals in the survival cohort to the County SES-LT by age, sex, race, calendar year and county of residence at the time of cancer diagnosis. SES is accounted for through the county of residence.
We selected patients diagnosed between 2000 and 2012 in the 18 SEER registries with any cancer, to calculate 5-year and 10-year expected survival, relative survival using the US-LT [9], relative survival using the new County SES-LT and cause-specific survival. In this paper, we report 10-year survival because it maximizes differences. We censored individuals when they reached age 99. We excluded patients diagnosed by autopsy or death certificate and those with no follow-up information (zero survival time).
Cause-specific survival uses cancer death as the endpoint and censors people dying of other causes, at the end of the study date or at attained age of 99 years, whichever comes first. Because of inherent ambiguities in determining the underlying cause of death (for example, a metastatic site reported as the cause of death rather than the original cancer site [7], SEER developed a cause-specific death classification algorithm [15, 24] to better code deaths related to the specific cancer. This algorithm uses causes of deaths that are likely to be related to the cancer or because of a cancer diagnosis. In the comparisons between relative and cause-specific survival we only included people with one primary malignant cancer, as cause of death is more likely to be misclassified for people diagnosed with multiple tumors. We did not report survival statistics when the number of patients at diagnosis were less than 50.
Results
Table 1 displays 2010 state population counts for each race and the percent of the population in each SES quintile. Note that DC does not have county subdivisions, Hawaii and Alaska counties are collapsed over state, and some states contain counties with 4 or fewer quintiles.
Fig 1 displays an example of the estimated County SES-LT in terms of log mortality for males and females by race in the state of California in 2010. The figure also displays the fit of observed log mortality rates for blacks and API in the high SES group (Q4-Q5). The patterns observed in California were in general similar to other states and show that: mortality is lower for API followed by Hispanics, whites, and blacks. Although not shown, AIAN have slightly lower mortality compared to blacks and there is a good fit of the County SES-LT to data. Figures for other states will be available in a website or by request.
Estimated and selected observed 2010 log mortality rates for males (A) and females (B) residing in the state of California by race and socio-economic status (SES). Observed log mortality rates are displayed for blacks and API in the high SES group (Q4-Q5). White = Non-Hispanic (NH) White, Black = NH Black, API = NH Asian or Pacific Islander, AIAN = NH American Indian and American Natives, and Hisp. = Hispanics. For whites only, LT for the lowest (Q1) and highest (Q5) SES quintiles are shown. For blacks, Hispanics, and API we only estimated LT by 2 level SES: Low = Q1-Q3 SES and High = Q4-Q5 SES. Observed log mortality rates for blacks and for API’s are for the high SES group.
Trends in life expectancies by race and SES
Between 1992 and 2012, male life expectancy increased more rapidly than female life expectancy in all races (Fig 2). Black men experienced the highest gains in life expectancy, 6.8 years, followed by Hispanic men (5.6 years), API men (5.0 years) and white men (3.8 years). Life expectancies among black, API, Hispanic and white women increased 4.4 years, 3.9 years, 2.8 years, and 2.0 years respectively. AIAN men and women experienced the smallest gains in life expectancy, 1.2 and 0 years, respectively. Gains in life expectancy were slightly higher for people living in counties with higher SES (Fig 2).
Life expectancies are from birth to age 99 and the numbers in parentheses represent gains in life expectancy (in years) from 1992 to 2012.
Life expectancies by state
Fig 3 displays linked (micro-)maps to facilitate geographical visualization of clusters of life expectancies by state. The dots represent the 2010 average life expectancy (in years) by state for white males, black males, white females, and black females. The horizontal bars represent the range of life expectancies between the lowest and highest SES quintile, the left and right end of the bar being the life expectancies for the lowest and highest SES quintile, respectively. The figure orders states by white male life expectancy from lowest (West Virginia, 73 years) to highest (DC, 82 years). The ordered states are color-coded in groups of 5 (panel 1), and the colors provide a link between the life expectancy estimates for the different race and sex (panels 2–5) and the maps (panel 6). Each map uses color to highlight states in that particular group. For example, the first map colors the states with the 5 lowest white male life expectancies. As we go down the maps, the next states with lower expectancy are colored and all the previous colored states (which have lower life expectancies) are displayed in grey. In the second map West Virginia, Mississippi, Kentucky, Alabama, and Arkansas are grey and Oklahoma, Tennessee, Louisiana, Nevada and South Carolina are colored. The maps show that south-eastern states have the lowest life expectancy and north-western states have the highest. There are significant differences in life expectancy by state, ranging from 73 to 82 years for white males, 78 to 86 years for white females, 66 to 75 for black males, and 75 to 81 for black females. States with wide bars show large disparities in life expectancy within the state, e.g. Idaho, Maryland, New Mexico, Tennessee and Virginia. Washington DC has the highest life expectancy for whites and the lowest for blacks.
The left and right end of the bars representing life expectancies estimates of patients living in counties with lowest and highest SES quintiles, respectively. The dots represent average life expectancy in the respective state, race and sex combination.
Impact of LT on expected survival and relative survival
Fig 4 displays 10-year expected survival for men and women diagnosed with any cancer by age and race using the County SES vs. the US-LT. Numbers can be interpreted as the percent surviving other causes of death 10 years after diagnosis estimated from each LT. Registries are ordered from higher SES to lower SES. In general, differences between expected survival from national and County SES-LT were small for whites, blacks and API’s and larger for the AIAN and Hispanics cancer patient’s populations. However, for whites and to a lesser extent for blacks there is consistent pattern between low vs. high SES areas. While the US-LT produced mostly flat expected survival the expected survival from the County SES life tables has a gradient with expected survival being higher in high SES areas (Hawaii, Connecticut, California, Seattle, Utah) and lower in low SES areas (Detroit, Georgia, Kentucky and Louisiana). The effect is higher for white, males and older ages 75–84. We are not able to observe this gradient for the County SES expected survival for AIAN, API and Hispanics because we were not able to estimate state specific LT for many states. The US and County SES expected survival were very similar for APIs with the largest differences occurring for males in Hawaii, and state for which there is state specific County SES-LT. The largest difference between US and County SES expected survival was observed for AIAN cancer population. The County SES expected survival is much lower, often 20% points lower than US expected survival, as the US-LT was available for other races, grouping AIAN with API mortality data. Because the API population is larger compared to the AIAN population, the US other race LT better reflect API’s than AIANs. The Hispanic expected survival is higher than the expected survival from US-LT, except for New Mexico.
Registries are grouped from the highest SES (left of SEER-18) to the lowest SES (right of SEER-18).
Fig 5 displays 10-year relative survival, using County SES-LT and US-LT and cause-specific survival for all cancer sites combined by sex, age and race. We chose to report 10-year survival as differences are maximized. Overall differences between the three survival measures were small. We highlight the main systematic differences. Among whites, especially men aged 75–84, the US-LT overestimates and underestimates relative survival compared to County SES-LT in high and low SES areas, respectively. Both County SES-LT relative survival and cause-specific survival show less of a gradient compared to US-LT relative survival, underscoring the fact that national average LT may increase differences in relative survival. Like expected survival, the largest differences between County SES-LT and US-LT relative survival were observed for AIAN and Hispanics cancer patients. For the AIAN cancer patients the County SES-LT relative survival is much higher, often higher than 10%, points, compared to relative survival from the US-LT and closer to cause-specific survival. For Hispanics, County SES-LT relative survival was lower relative to the US-LT relative survival especially for cancer patients aged 75–84 and closer to cause-specific survival.
Registries are grouped from the highest SES (left of SEER-18) to the lowest SES (right of SEER-18).
Table 2 presents a summary of the comparisons between 10-year relative survival using US-LT, County SES-LT and 10-year cause specific survival for cancer patients diagnosed in the aggregated SEER-18 areas between ages 75–84 years by race. Except for whites, 10-year County SES-LT relative survival was closer to 10-year cause-specific survival. The largest differences were seen among AIAN and Hispanics. Relative survival using County SES-LT was 7.3% and 6.7% survival points closer to cause-specific survival compared to US-LT relative survival for AIAN and Hispanics cancer patients, respectively.
Discussion
In this study, we developed an extensive set of life tables (LT) representing mortality patterns in the United States over three decades by race, ethnicity, sex, geography, and county-level SES. These tables can help calculate life expectancy and improve relative survival estimates. Despite gains in life expectancy between 1992 and 2012, this study shows that large disparities in life expectancy continue to exist among sex, race, state, and socio-economic (SES) groups. Asian or Pacific Islanders (API) had the highest life expectancy, followed by Hispanics, whites, blacks, and American Indian and Alaska Natives (AIAN). Black and AIAN men had the lowest life expectancy [25]. Between 1992 and 2012, the largest increases in life expectancy were observed among black men [26], and no increase was observed for AIAN women. Our findings suggest that differences in race, geography, and SES have a greater effect on life expectancy and on County SES-LT relative survival among males compared to females. Previous research has shown that most variation in life expectancy is due to differences in health behaviors, including smoking and obesity [12]. Thus the larger impact of race, geography, and SES on males life expectancy may be attributed to the fact that males have a higher and larger geographical variability in smoking prevalence compared to females [27].
The main use of these life tables is for the reporting of relative survival from U.S. cancer registries. Differences between relative survival calculated using the County SES-LT versus the US-LT were in general small, particularly for younger cancer patients, for areas with SES comparable to the national average (e.g. SEER-18), and for survival of 5-years or less (data not shown). Differences were larger for older ages, whites in high or low SES areas, AIANs, Hispanics, APIs in Hawaii, and long term survival. Compared to US-LT, the new County SES-LT provide lower expected survival in lower SES areas and higher expected survival in higher SES areas. Consequently, relative survival using the County SES-LT is lower in areas with high SES and higher in areas with low SES, decreasing differences in relative survival. This was more evident for whites and for males, as LT were more detailed and included both state and the full 5 levels of SES at the county level. Specific AIAN LT improved and increased estimates of AIAN relative survival by more than 10% survival points in most cases. The County SES-LT relative survival estimates were in general closer to cause-specific survival than US-LT relative survival, demonstrating that County SES-LT better captured background mortality especially in high versus low SES areas, male patients diagnosed at older ages, and among AIAN and Hispanics.
Comparisons between relative survival and cause-specific survival are challenging[15, 28], since both can be subject to bias. We used all cancer sites and included only cancer patients with one tumor to improve comparability between relative versus cause-specific survival. This comparison also shows that although the County SES-LT were closer to cause-specific survival, there were still some systematic differences. For example, for white males aged 75–84, the County SES-LT relative survival was parallel but higher than cause-specific survival by an average 7% points. The large percent of prostate cancers (30%) diagnosed among men, compared to all other cancers, may explain this difference along with the healthy screening effect. The healthy screening effect postulates that people detected with cancer through screening may have a higher life expectancy than the general US population, perhaps because of better overall health, greater access to health care, or healthier lifestyles. The healthy screening effect was most recently demonstrated among prostate cancer patients in the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial. Participants in this trial had a 30–50% lower mortality rate for heart disease, injury, and kidney disease than expected in the general population[29].
There were limitations to our study. We used mortality data linked to an SES indicator at the county level, which is subject to large variability in population size. For large counties, such as Los Angeles County, California, with a population close to 10 million, the SES index represents the average of SES in the county for all residents and does not have the required specificity to characterize the diversity of SES in the county. Analysis at the census tract level would be more attractive, since census tract populations are more homogeneous, however county is the smallest geographical level for which U.S. mortality data are available. The composite SES index summarizes a full range of SES characteristics to simplify the modeling but still accounts for variation in SES level with geography and race. The quintiles derived are relative measures of SES in the U.S., and useful for comparing inequalities between counties. The quintiles also showed a consistent pattern, in which higher life expectancies were associated with higher SES quintiles and vice versa.
Importantly, modeling separately for each race at the individual level already considers important SES differences. For example, DC represented by one county (Q4), provided the highest life expectancy for white men (82 years) and the lowest for black men (66 years). Second, because of a small number of deaths, we varied the models with respect to geography and SES levels by race groups and restricted analyses to ages 30 to 84 to provide robust LT estimates. Although this approach may not have captured all the variability in the mortality in the older and younger age groups, restricting the modeling to ages 30 to 84 and borrowing information from the national NCHS LT provided more stable and less biased estimates of LT in those age groups. A previous study has shown that state life tables using mortality data beyond ages 85 provided unreliable estimates of relative survival [11]. The NCHS national life tables are more reliable for older age groups because they use Medicare data, not available to us, to provide a more accurate determination of age of death for older individuals. Because mortality at younger ages is very low, the impact of life tables on relative survival from cancer patients diagnosed at younger ages is very small. Our estimates based on national LT are more robust and not subject to potential biases due to data variability at the younger ages when mortality is small.
We also restricted estimation of the AIAN LT to CHSDA areas, similar to other studies [17]. Life expectancy for the AIAN population in CHSDA areas, which are predominantly rural, isolated areas with limited access to employment and health care, may not well represent the total AIAN populations. However, estimates not restricting to CHSDA areas would result in unrealistically high estimates of AIAN life expectancy. Our estimates are similar to Arias et al.[17], a life expectancy of 68.3 in 2010 versus a life expectancy of 68 years in 2007–2009 for non-Hispanic AIAN in Arias et al. [17].
The County SES-LT included calendar year, age, sex, race, area of residency and County SES when possible. We were not able to include other variables at the county level, such as risk factors that may affect other causes of death, e.g. smoking or obesity, or variables related to access to health care, as these data are not available for the full range of years and for all counties. Previous studies have shown that LT adjusting for higher, smoking-related background mortality, had little or modest impact on relative survival estimates for lung cancer [30, 31].
Strengths of our study include the large sample size, the population-based setting, and the fact that the LT are an extensive representation of the varying mortality patterns in the U.S. over three decades by race, ethnicity, sex, geography and county-level SES. Our study highlighted the importance of LT by geography and other factors for comparisons and calculation of relative survival among different cancer registries, due to the disparities seen in life expectancy across different subgroups in the U.S. Analyses of life expectancies from other studies provided comparable results[12, 25–26], giving validity to these LTs. The comparisons of relative and cause-specific survival showed that the County SES relative survival were closer to cause-specific survival and had a smaller gradient between low versus high SES areas, reducing differences in relative survival. This substantiated the fact that relative survival using a national average background mortality (thus same denominator) overestimates and underestimates survival in high and low SES areas, respectively.
In summary, we have shown that differences between relative survival using the SES and the US-LT were in general small. However, relative survival using County SES-LT were closer to cause-specific survival and improved estimates for some demographic groups, in particular Hispanics, AIAN, populations in higher or lower SES areas/states, and among older male cancer patients. Investigations using SEER data to examine time trends in cancer survival and disparities in cancer survival by race, ethnicity, and SES are common. Studies of cancer survival using life tables that do not properly account for differences in background mortality by these factors may mischaracterize trends and overstate the magnitude of disparities. Recently, the North American Association of Central Cancer Registries (NAACCR) began to publish cancer survival estimates on a larger number of U.S. state registries in the Cancer in North America annual reports [13]. We suggest using the life tables described in this paper as default for cancer relative survival using U.S. data, including the CDC’s National Program of Cancer Registries, the SEER registries, and by researchers conducting international studies that include U.S. data [14].The use of these life tables will advance population-based cancer surveillance and research by contributing standardized and more accurate relative survival estimates.
Supporting information
S1 Table. Life table model description for each state and race.
https://doi.org/10.1371/journal.pone.0201034.s001
(XLSX)
References
- 1. Arias E, Eschbach K, Schauman WS, Backlund EL, Sorlie PD. The Hispanic mortality advantage and ethnic misclassification on US death certificates. Am J Public Health. 2010;100 Suppl 1:S171–7.
- 2. Wei R, Anderson RN, Curtin LR, Arias E. U.S. decennial life tables for 1999–2001: state life tables. Natl Vital Stat Rep. 2012;60(9):1–66. pmid:24979971
- 3.
Arias EH, M.; Xu, J. United States Life Tables, 2014. National Vital Statistics Reports 2017;66: 4 [Internet]. [Cited October 3, 2017]. Available from: https://www.cdc.gov/nchs/data/nvsr/nvsr66/nvsr66_04.pdf.
- 4. Mariotto AB, Noone AM, Howlader N, Cho H, Keel GE, Garshell J, et al. Cancer survival: an overview of measures, uses, and interpretation. Journal of the National Cancer Institute Monographs. 2014;2014(49):145–86. pmid:25417231
- 5. Ederer F, Axtell LM, Cutler SJ. The relative survival rate: a statistical methodology. NatlCancer InstMonogr. 1961;6:101–21.
- 6. Baili P, Micheli A, De Angelis R, Weir HK, Francisci S, Santaquilani M, et al. Life tables for world-wide comparison of relative survival for cancer (CONCORD study). Tumori. 2008;94(5):658–68. pmid:19112937
- 7. Percy C, Stanek E 3rd, Gloeckler L. Accuracy of cancer death certificates and its effect on cancer mortality statistics. Am J Public Health. 1981;71(3):242–50. pmid:7468855
- 8.
Noone AM, Howlader N, Krapcho M, Miller D, Brest A, Yu M, Ruhl J, et al. (eds). SEER Cancer Statistics Review, 1975–2015, National Cancer Institute. Bethesda, MD, based on November 2017 SEER data submission, posted to the SEER web site, April 2018. [Cited 17 May 2018]. Available from: https://seer.cancer.gov/csr/1975_2015/.
- 9.
Expected Survival Life Tables. [Cited 20 November 2017]. In National Cancer Institute SEER [Internet]. Available from https://seer.cancer.gov/expsurvival/.
- 10. Azzani M, Roslani AC, Su TT. The perceived cancer-related financial hardship among patients and their families: a systematic review. Support Care Cancer. 2015;23(3):889–98. pmid:25337681
- 11. Stroup AM, Cho H, Scoppa SM, Weir HK, Mariotto AB. The impact of state-specific life tables on relative survival. Journal of the National Cancer Institute Monographs. 2014;2014(49):218–27. pmid:25417235
- 12. Chetty R, Stepner M, Abraham S, Lin S, Scuderi B, Turner N, et al. The Association Between Income and Life Expectancy in the United States, 2001–2014. JAMA: the journal of the American Medical Association. 2016;315(16):1750–66. pmid:27063997
- 13.
Johnson CJ, Mariotto A, Nishri D, Weir HK,Wilson R, Copeland G, et al. (eds). Cancer in North America: 2010–2014 Volume Four: Cancer Survival in the United States and Canada 2007–2013. Springfield, IL: North American Association of Central Cancer Registries, Inc. June 2017. [Cited 2017 Nov 20]. Available from: https://www.naaccr.org/cancer-in-north-america-cina-volumes/.
- 14. Allemani C, Weir HK, Carreira H, Harewood R, Spika D, Wang XS, et al. Global surveillance of cancer survival 1995–2009: analysis of individual data for 25,676,887 patients from 279 population-based registries in 67 countries (CONCORD-2). Lancet. 2015;385(9972):977–1010. pmid:25467588
- 15. Howlader N, Ries LA, Mariotto AB, Reichman ME, Ruhl J, Cronin KA. Improved estimates of cancer-specific survival rates from population-based data. J Natl Cancer Inst. 2010;102(20):1584–98. pmid:20937991
- 16. Mariotto AB, Yabroff KR, Shao YW, Feuer EJ, Brown ML. Projections of the Cost of Cancer Care in the United States: 2010–2020. Journal of the National Cancer Institute. 2011;103(2).
- 17. Arias E, Xu JQ, Jim MA. Period Life Tables for the Non-Hispanic American Indian and Alaska Native Population, 2007–2009. Am J Public Health. 2014;104:S312–S9. pmid:24754553
- 18. Yu MD, Tatalovich Z, Gibson JT, Cronin KA. Using a composite index of socioeconomic status to investigate health disparities while protecting the confidentiality of cancer registry data. Cancer Cause Control. 2014;25(1):81–92.
- 19. Arias E, Curtin LR, Wei R, Anderson RN. U.S. decennial life tables for 1999–2001, United States life tables. NatlVital StatRep. 2008;57(1):1–36.
- 20.
Elandt-Johnson RC, Johnson NL. Survival Models and Data Analysis: Wiley; 1990.
- 21.
SEER*Stat software. [Cited 20 November 2017]. In National Cancer Institute, Surveillance, Epidemiology, and End Results (SEER) Data and Software [Internet]. Available from: www.seer.cancer.gov/seerstat.
- 22.
Ederer F, Heise H. Intructions to IBM 650 programmers in processing survival computations. Methodological note no. 10, end results evaluation section. Technical report,. National Cancer Insitute, Bethesda, MD, 1959.
- 23.
Cho H HN, Mariotto A, Cronin K. Estimating relative survival for cancer patients from the SEER Program using expected rates based on Ederer I versus Ederer II method. Surveillance Research Program, NCI, Technical Report #2011–01. 2011. Available from: http://surveillancecancergov/reports/tech201101pdf. Cited 20 November 2017.
- 24. Verdecchia A, Capocaccia R, Egidi V, Golini A. A method for the estimation of chronic disease morbidity and trends from mortality data. Stat Med. 1989;8(2):201–16. pmid:2784863
- 25. Olshansky SJ, Antonucci T, Berkman L, Binstock RH, Boersch-Supan A, Cacioppo JT, et al. Differences in life expectancy due to race and educational differences are widening, and many may not catch up. Health Aff (Millwood). 2012;31(8):1803–13.
- 26. Fuchs VR. Black Gains in Life Expectancy. JAMA: the journal of the American Medical Association. 2016;316(18):1869–70. pmid:27656867
- 27. Jamal A, King BA, Neff LJ, Whitmill J, Babb SD, Graffunder CM. Current Cigarette Smoking Among Adults—United States, 2005–2015. Mmwr-Morbid Mortal W. 2016;65(44):1205–11.
- 28. Skyrud KD, Bray F, Moller B. A comparison of relative and cause-specific survival by cancer site, age and time since diagnosis. Int J Cancer. 2014;135(1):196–203. pmid:24302538
- 29. Pinsky PF, Miller A, Kramer BS, Church T, Reding D, Prorok P, et al. Evidence of a healthy volunteer effect in the prostate, lung, colorectal, and ovarian cancer screening trial. American journal of epidemiology. 2007;165(8):874–81. pmid:17244633
- 30. Hinchliffe SR, Dickman PW, Lambert PC. Adjusting for the proportion of cancer deaths in the general population when using relative survival: a sensitivity analysis. Cancer Epidemiol. 2012;36(2):148–52. pmid:22000329
- 31. Ellis L, Coleman MP, Rachet B. The impact of life tables adjusted for smoking on the socio-economic difference in net survival for laryngeal and lung cancer. British journal of cancer. 2014;111(1):195–202. pmid:24853177