Trends and Patterns of Disparities in Cancer Mortality Among
US Counties, 1980–2014
Ali H. Mokdad, PhD, Laura Dwyer-Lindgren, MPH, Christina Fitzmaurice, MD, MPH,
Rebecca W. Stubbs, BA, Amelia Bertozzi-Villa, MPH, Chloe Morozoff, MPH, Raghid
Charara, MD, Christine Allen, BA, Mohsen Naghavi, MD, PhD, and Christopher J. L. Murray,
MD, DPhil
Institute for Health Metrics and Evaluation, University of Washington, Seattle
Abstract
INTRODUCTION—Cancer is a leading cause of morbidity and mortality in the United States and
results in a high economic burden.
OBJECTIVE—To estimate age-standardized mortality rates by US county from 29 cancers.
DESIGN AND SETTING—Deidentified death records from the National Center for Health
Statistics (NCHS) and population counts from the Census Bureau, the NCHS, and the Human
Mortality Database from 1980 to 2014 were used. Validated small area estimation models were
used to estimate county-level mortality rates from 29 cancers: lip and oral cavity; nasopharynx;
other pharynx; esophageal; stomach; colon and rectum; liver; gallbladder and biliary; pancreatic;
larynx; tracheal, bronchus, and lung; malignant skin melanoma; nonmelanoma skin cancer; breast;
cervical; uterine; ovarian; prostate; testicular; kidney; bladder; brain and nervous system; thyroid;
mesothelioma; Hodgkin lymphoma; non-Hodgkin lymphoma; multiple myeloma; leukemia; and
all other cancers combined.
EXPOSURE—County of residence.
MAIN OUTCOMES AND MEASURES—Age-standardized cancer mortality rates by county,
year, sex, and cancer type.
Corresponding Author: Christopher J. L. Murray, MD, DPhil, Institute for Health Metrics and Evaluation, 2301 Fifth Ave, Ste 600,
Seattle, WA 98121 (cjlm@uw.edu).
Conflict of Interest Disclosures: All authors have completed and submitted the ICMJE Form for Disclosure of Potential Conflicts of
Interest and none were reported.
Author Contributions: Dr Murray had full access to all of the data in the study and takes responsibility for the integrity of the data
and the accuracy of the data analysis.
Concept and design:
Mokdad, Dwyer-Lindgren, Bertozzi-Villa, Charara, Naghavi, Murray.
Acquisition, analysis, or interpretation of data:
Mokdad, Dwyer-Lindgren, Fitzmaurice, Stubbs, Bertozzi-Villa, Morozoff, Allen,
Naghavi.
Drafting of the manuscript:
Mokdad, Dwyer-Lindgren, Charara, Allen.
Critical revision of the manuscript for important intellectual content:
Mokdad, Dwyer-Lindgren, Fitzmaurice, Stubbs, Bertozzi-Villa,
Morozoff, Naghavi, Murray.
Statistical analysis:
Mokdad, Dwyer-Lindgren, Stubbs, Bertozzi-Villa, Allen, Naghavi.
Obtained funding:
Mokdad, Murray.
Administrative, technical, or material support:
Mokdad, Morozoff, Allen, Murray.
Supervision:
Murray.
HHS Public Access
Author manuscript
JAMA
. Author manuscript; available in PMC 2017 September 27.
Published in final edited form as:
JAMA
. 2017 January 24; 317(4): 388–406. doi:10.1001/jama.2016.20324.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
RESULTS—A total of 19 511 910 cancer deaths were recorded in the United States between
1980 and 2014, including 5 656 423 due to tracheal, bronchus, and lung cancer; 2 484 476 due to
colon and rectum cancer; 1 573 593 due to breast cancer; 1 077 030 due to prostate cancer; 1 157
878 due to pancreatic cancer; 209 314 due to uterine cancer; 421 628 due to kidney cancer; 487
518 due to liver cancer; 13 927 due to testicular cancer; and 829 396 due to non-Hodgkin
lymphoma. Cancer mortality decreased by 20.1%(95% uncertainty interval [UI], 18.2%–21.4%)
between 1980 and 2014, from 240.2 (95% UI, 235.8–244.1) to 192.0 (95% UI, 188.6–197.7)
deaths per 100 000 population. There were large differences in the mortality rate among counties
throughout the period: in 1980, cancer mortality ranged from 130.6 (95% UI, 114.7–146.0) per
100 000 population in Summit County, Colorado, to 386.9 (95% UI, 330.5–450.7) in North Slope
Borough, Alaska, and in 2014 from 70.7 (95% UI, 63.2–79.0) in Summit County, Colorado, to
503.1 (95% UI, 464.9–545.4) in Union County, Florida. For many cancers, there were distinct
clusters of counties with especially high mortality. The location of these clusters varied by type of
cancer and were spread in different regions of the United States. Clusters of breast cancer were
present in the southern belt and along the Mississippi River, while liver cancer was high along the
Texas-Mexico border, and clusters of kidney cancer were observed in North and South Dakota and
counties in West Virginia, Ohio, Indiana, Louisiana, Oklahoma, Texas, Alaska, and Illinois.
CONCLUSIONS AND RELEVANCE—Cancer mortality declined overall in the United States
between 1980 and 2014. Over this same period, there were important changes in trends, patterns,
and differences in cancer mortality among US counties. These patterns may inform further
research into improving prevention and treatment.
Cancer is the second leading cause of death in the United States and globally.
1
Moreover,
cancer is a major cause of morbidity in the United States
1
and is associated with a high
economic burden.
2
Overall cancer mortality rates have declined in the United States in
recent decades; however, major differences in cancer mortality still exist.
3
Several studies have reported on the variation in cancer mortality by state.
4,5
This variation
is at least partially explained by differences in risk factors, socioeconomic factors, and
access to high-quality treatment.
6
For example, smoking rates have declined in the United
States, but this decline varied by location.
7
Similarly, while obesity increased in recent years
throughout the United States,
8
the rate of increase varied widely.
9
Moreover, access to health
care and the quality of available health care varies tremendously among states and different
socioeconomic groups.
10
Most previous reports on geographic differences in cancer mortality have focused on
variation by state, with less information available at the county level.
4
There is a value for
data at the county level because public health programs and policies are mainly designed and
implemented at the local level. Moreover, local information can also be useful for health
care clinicians to understand community needs for care and aid in identifying cancer hot
spots that need more investigation to understand the root causes.
Methods
This analysis used methods reported in detail elsewhere.
11
A brief description of this
approach and its application to cancer mortality is provided below. This research received
Mokdad et al. Page 2
JAMA
. Author manuscript; available in PMC 2017 September 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
institutional review board approval from the University of Washington. Informed consent
was not required because the study used deidentified data and was retrospective.
Data
Deidentified death records from the National Center for Health Statistics (NCHS)
12
and
population counts from the Census Bureau,
13
the NCHS,
14–16
and the Human Mortality
Database were used.
17
Deaths and population were tabulated by county, age group (0, 1–4,
5–9, …, 75–79, and ≥80 years), sex, year, and (in the case of death data) cause. County-level
information on levels of education, income, race/ethnicity, Native American reservations,
and population density derived from data provided by the Census Bureau and the NCHS
were also used. More detail on these data sources is provided in eTable 1 in the Supplement.
Cause List and Garbage Code Redistribution Methods
The study used the cause list developed for the Global Burden of Diseases, Injuries, and
Risk Factors Study (GBD).
1
This cause list is arranged hierarchically in 4 levels, and within
each level the list is exhaustive and mutually exclusive. eTable 2 in the Supplement lists all
causes in the GBD cause list and the
International Classification of Diseases, Ninth Revision
and
International Statistical Classification of Diseases and Related Health Problems, Tenth
Revision
codes that correspond to each cause. Although the focus of this study is cancers, all
causes of death in the GBD cause list were analyzed concurrently.
Previous studies
1
have documented the existence of “garbage codes” in death registration
data, which may lead to misleading spatial and temporal patterns, as well as misleading
ranks among causes, as the percentage of deaths assigned garbage codes varies by location,
year, and true underlying cause. This study used garbage redistribution methods developed
for the GBD to reallocate deaths assigned garbage codes.
1
First, plausible target causes were
identified for each garbage code or group of garbage codes. Second, deaths were reassigned
to the specified target codes according to proportions derived in 1 of 4 ways: (1) published
literature or expert opinion; (2) regression models; (3) according to the proportions initially
observed among targets; and (4) for HIV/AIDS specifically, by comparison with years
before HIV/AIDS became widespread.
Small Area Models
The study estimated spatially explicit Bayesian mixed-effects regression models for cancer
mortality in the GBD hierarchy, separately for males and females. The model for each cause
was specified as
where
D
j,t,a
,
P
j,t,a
, and
m
j,t,a
are the number of deaths, the population, and the underlying
mortality rate, respectively, for county
j
, year
t
, and age group
a.
The model for
m
j,t,a
contained 6 components: an intercept (β
0
), fixed covariate effects (β
1
),random age-time
effects (γ
1,
a
,
t
),random spatial effects (γ
2,
j
), random space-time effects (γ
3,
j
and γ
4,
j
,
t
), and
Mokdad et al. Page 3
JAMA
. Author manuscript; available in PMC 2017 September 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
random space-age effects (γ
5,
j
and γ
6,
j
,
a
). The model incorporated 7 covariates: the
proportion of the adult population that graduated high school, the proportion of the
population that is Hispanic, the proportion of the population that is black, the proportion of
the population that is a race other than black or white, the proportion of a county that is
contained within a state or federal Native American reservation, the median household
income, and the population density. γ
1
, γ
2
, γ
3
, and γ
5
were assumed to follow conditional
autoregressive distributions, which allow for smoothing over adjacent age groups and years
(γ
1
) or counties (γ
2
, γ
3
, and γ
5
). γ
4
and γ
6
were assumed to follow independent mean-zero
normal distributions.
Models were fit using the Template Model Builder Package
18
in R version 3.2.4.
19
Model
predictions were then raked (ie, iteratively scaled along multiple dimensions) to ensure
consistency between levels of the cause hierarchy and simultaneously ensure consistency
with existing national-level estimates from the GBD. After raking, age-standardized
mortality rates were calculated using the US 2010 Census population as the standard, and
years of life lost were calculated by multiplying the mortality rate by population by life
expectancy at the average age at death in the reference life table used in the GBD
1
and then
summing across all ages. When measuring changes over time, the change was considered
statistically significant if the posterior probability of an increase (or decrease) was at least
95%. No explicit correction for multiple testing (ie, across multiple counties) was applied;
however, modeling all counties simultaneously is expected to mitigate the risk of spuriously
detecting changes due to multiple testing. The study reports mortality rates for lip and oral
cavity; nasopharynx; other pharynx; esophageal; stomach; colon and rectum; liver;
gallbladder and biliary; pancreatic; larynx; tracheal, bronchus, and lung; malignant skin
melanoma; nonmelanoma skin cancer; breast; cervical; uterine; ovarian; prostate; testicular;
kidney; bladder; brain and nervous system; thyroid; mesothelioma; Hodgkin lymphoma;
non-Hodgkin lymphoma; multiple myeloma; leukemia; and all other cancers combined.
Risk Factors and Population-Attributable Fraction
Full detailed methods for calculating relative risks and population-attributable fractions are
available elsewhere.
20
Briefly, GBD 2015 used the comparative risk assessment framework
developed for previous iterations of the GBD study to estimate attributable deaths,
disability-adjusted life-years, and trends in exposure by age group, sex, year, and geography
for 79 behavioral, environmental and occupational, and metabolic risks or clusters of risks
over the period 1990 to 2015. Two types of risk assessments are possible within the
comparative risk assessment framework: attributable burden and avoidable burden.
Attributable burden is the reduction in current disease burden that would have been possible
if past population exposure had shifted to an alternative or counterfactual distribution of risk
exposure. Avoidable burden is the potential reduction in future disease burden that could be
achieved by changing the current distribution of exposure to a counterfactual distribution of
exposure. Four types of counterfactual exposure distributions have been identified
21
: (1)
theoretical minimum risk; (2) plausible minimum risk; (3) feasible minimum risk; and (4)
cost-effective minimum risk. In GBD studies and in this study, the focus was on attributable
burden using the theoretical minimum risk level, which is the level of risk exposure that
Mokdad et al. Page 4
JAMA
. Author manuscript; available in PMC 2017 September 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
minimizes risk at the population level or the level of risk that captures the maximum
attributable burden.
Results
Table 1 shows the number of deaths, years of life lost, and age-standardized mortality rates
at the national level as well as the distribution of age-standardized mortality rates at the
county level for each cancer. The study reported the number of years of life lost in addition
to deaths to account for the fact that many deaths from certain cancers occur at an older age.
For example, prostate cancer was the fifth leading cause of death among cancers but the
ninth leading cause of cancer years of life lost. Lung, colon, and breast cancer were the top 3
leading causes for all metrics. Lung, colon, and breast cancers also had the largest absolute
difference in mortality between counties at the 90th percentile and the 10th percentile. Lung
cancer mortality rates were twice as high among counties in the 90th percentile compared
with counties in the 10th.
Table 2 shows the 5-year relative survival for selected cancers from the Surveillance,
Epidemiology, and End Results program
22,23
(the corresponding age-specific estimates are
given in eTable 3 in the Supplement) and the population-attributable fraction from the GBD
using the comparative risk assessment approach.
20
Although cancer survival improved from
1973 to 2014 for all cancers, 6 cancers (testicular, thyroid, prostate, breast, melanoma, and
Hodgkin lymphoma)had a 5-year survival rate of more than 85%. The population
attributable fraction of risk factors was the highest for lung and cervical cancer and the
lowest for ovarian cancer.
Results for all cancers combined and for 10 specific cancers are presented below and
graphically in the Figures, with results for the remaining cancers presented in eFigures 1–23
in the Supplement. The 10 specific cancers highlighted below were chosen because they
have either high burden (eg, tracheal, bronchus, and lung cancer), because they are
responsive to treatment (eg, testicular cancer), or because screening is an important
component of the health system response (eg, breast cancer). For cancers that predominantly
or exclusively affect males or females (eg, breast cancer, prostate cancer), results are
reported for males or females only, while in all other cases results are presented for both
sexes combined. Mortality rates by county for each cancer are available in an online
visualization tool (Interactive).
Mokdad et al. Page 5
JAMA
. Author manuscript; available in PMC 2017 September 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript