Background Disparities in cancer incidence and mortality have been observed by measures of area-level socioeconomic status (SES); however, the extent to which these disparities are explained by individual SES is unclear.
Methods Participants included 60 756 men and women in the VITamins And Lifestyle (VITAL) study cohort, aged 50–76 years at baseline (2000–2002) and followed through 2010. We constructed a block group SES index using the 2000 US Census and fit Cox proportional hazards models to estimate the association between area-level SES (by quintile) and total and site-specific cancer incidence and total cancer mortality, with and without household income and individual education in the models.
Results Lower area-level SES was weakly associated with higher total cancer incidence and lower prostate cancer risk, but was not associated with risk of breast cancer. Compared with the highest-SES areas, living in the lowest-SES areas was associated with higher lung (HR: 2.21, 95% CI 1.69 to 2.90) and colorectal cancer incidence (HR: 1.52, 95% CI 1.11 to 2.09) and total cancer mortality (HR: 1.68, 95% CI 1.47 to 1.93). Controlling for individual education and household income weakened the observed associations, but did not eliminate them (lung cancer HR: 1.43, 95% CI 1.07 to 1.91; colorectal cancer HR: 1.35, 95% CI 0.97 to 1.88; cancer mortality HR: 1.28, 95% CI 1.11 to 1.48).
Conclusions Area-level socioeconomic disparities exist for several cancer outcomes. These differences are not fully explained by individual SES, suggesting area-level factors may play a role.
- Cohort studies
- Health inequalities
Statistics from Altmetric.com
Associations have been reported between area-level socioeconomic status (SES) and several cancer outcomes—lower area-level SES has been associated with higher risk of colorectal,1–4 lung,3 ,5 prostate6 and cervical cancer;7 ,8 total9 ,10 and site-specific cancer mortality;10 ,11 later stage of diagnosis;12–18 and more aggressive tumour characteristics19; while higher area-level SES has been associated with higher risk of breast8 ,20–22 and prostate cancer5 ,8 ,23—however, the extent to which these observed associations are due to individual SES is rarely addressed.6
Understanding the extent to which observed associations between area-level SES and cancer outcomes are due to compositional factors (eg, if people living in lower-SES areas are themselves of lower SES and would be at increased risk of disease and mortality regardless of where they lived), or potentially influenced by contextual factors (eg, physical environment, neighbourhood resources, policies or social norms, which may contribute to disease risk independent of individual SES) is critical for appropriately targeting interventions to reduce socioeconomic disparities.16 ,24
The purpose of this paper is to estimate the association between area-level SES and total and site-specific cancer incidence and total cancer mortality, and to assess whether observed associations remain after control for individual educational attainment and household income. While the first approach estimates total area-level socioeconomic disparities in cancer outcomes, the second evaluates the degree of disparity that could be due to contextual effects of areas on cancer outcomes or their risk factors. To the best of our knowledge this is the first study to systematically examine whether observed associations between area-level SES and several cancer outcomes is due to individual socioeconomic characteristics by directly comparing those associations with and without control for individual SES.
Materials and methods
The VITamins And Lifestyle (VITAL) study is a prospective cohort study designed to investigate the associations of use of dietary supplements and other behaviours with cancer risk and mortality. It has previously been described in detail.25 Participants were between ages 50 and 76 and lived in 1 of the 13 counties in the Western Washington Surveillance, Epidemiology and End Results (SEER) cancer registry at baseline. The Institutional Review Board at the Fred Hutchinson Cancer Research Center approved this research.
Using names purchased from a commercial mailing list, 364 418 sex-specific baseline questionnaires were mailed between October 2000 and December 2002 and were followed 2 weeks later by reminder postcards. A total of 79 300 questionnaires were returned, of which 77 719 passed quality control checks. Overall, 60 756 men and women were included in the cancer mortality analyses after excluding respondents whose baseline addresses were post office boxes (n=1137) or could not be geocoded (n=381) and respondents missing data on education (n=1333) or household income (n=15 443). Missing individual education and household income were not associated with area-level SES. Models of area-level SES, and total and site-specific cancer incidence further excluded respondents with a history of cancer other than non-melanoma skin cancer (n=11 259) or whose history of cancer was unknown (n=214). Numbers of exclusions reported are not mutually exclusive.
Respondents’ baseline addresses were geocoded using GPS Visualizer and Yahoo Maps. A 1% sample of addresses was geocoded again using Google Maps and more than 95% of the addresses in the validation sample were geocoded to within 400 m of one another using the two methods. Addresses were used to identify respondents’ census block groups using TIGER/Line shapefiles for the 2000 US Census in ArcMap 10 (Esri, Redlands, California, USA).
Area-level SES was measured using a method developed by Diez-Roux et al26 that has been used previously to examine associations between area-level SES and colon and rectal cancer.2 Information from the 2000 Census was used to create a block group-level index of social disadvantage including log of median value of owner-occupied housing units; log of median household income; per cent of households receiving net rental, interest or dividend income; per cent of adults ages 25 and older who completed high school and who completed college; and per cent of employed persons ages 16 and older in professional and managerial occupations. Standardised z-scores were calculated for each variable based on the 3347 block groups in the Western Washington SEER catchment area and summed. Signs of the index scores were reversed so that higher values corresponded with lower area-level SES. Each participant was assigned the index value for their block group of residence. Index values ranged from −16.1 to 17.3 with a median value of −1.1 and mean of −1.3.
Block groups were chosen as an approximation of participants’ neighbourhood environments because they are small, relatively permanent statistical subdivisions of counties and of census tracts designed to be relatively homogenous with respect to population characteristics, economic factors and living conditions,27 and have been found to perform favourably in detecting socioeconomic gradients in cancer incidence and mortality.8 Block groups in the catchment area covered a median of 3.3 square miles and included a median population of 1070.
Case ascertainment and censoring
In cancer incidence analyses, participants with no history of cancer at baseline were followed for their first incident, invasive cancer via annual linkage with SEER. This linkage is largely automated and based on ranking agreement between items common to both sets of data, such as Social Security number, name and date of birth. Matches with high concordance were linked automatically whereas visual inspection was used to adjudicate incomplete matches. A total of 6099 incident cancers were identified in an average of 8.1 years of follow-up.
Participants not diagnosed with cancer were right-censored at the date of the earliest of the following events: date they requested removal from the study (n=8), date they moved out of the SEER catchment area (n=3898), date of death (n=2214) or 31 December 2010 (n=39 967). Moves out of area were identified through linkage with the US National Change of Address System. For analyses of site-specific cancer incidence, participants diagnosed with cancers other than the one of interest were censored at the date of cancer diagnosis.
Cancer deaths were ascertained through annual linkage with the Washington State death file using procedures similar to those described above. In cancer mortality analyses, participants who did not die of cancer were right-censored at the date they requested removal from the study (n=9), date they moved out of Washington State (n=3536), date of death due to other causes (n=3116) or 31 December 2010 (n=51 608). A total of 2487 cancer deaths were observed in an average of 8.5 years of follow-up.
Area-level SES was divided into quintiles based on the distribution of participants’ block group SES index values. Using these categories, cancer incidence and mortality rates were calculated and Cox proportional hazards models were used to calculate HRs and 95% CIs of cancer incidence and cancer mortality associated with living in areas in each of the lowest four quintiles of area-level SES compared with living in the highest-SES quintile. Participant age was used as the time scale, with participants entering the analysis at their age at baseline and exiting at age at outcome (cancer diagnosis; death due to cancer) or censoring event, as described above. Proportional hazards assumptions were examined using scaled Schoenfeld residuals. No significant deviations from proportionality were observed. All statistical tests were two sided with p<0.05 considered statistically significant.
Multivariable analyses included categorical variable adjustment for sex; additional adjustment for race/ethnicity and marital status (model 1) and further adjustment for individual education and household income (model 2). Although race/ethnicity and marital status are related to individual SES, the age and sex-adjusted model and model 1 yielded similar results for all cancer outcomes. p Values for trend are from the Wald test associated with area-level SES index modelled as a continuous variable. Statistical analyses were conducted using Stata 12.1 (StataCorp LP, College Station, Texas, USA). All models of area-level SES and cancer incidence and mortality utilise the cluster option to obtain SEs that account for correlation among residents of the same block groups.
Results of multilevel survival, or frailty, models with Weibull-distributed event times, γ frailty distributions and shared frailties by block group of residence are provided in online supplementary tables. These models use time since baseline as the time scale and include adjustment for baseline age. Results are nearly identical to the models presented here; however, not all frailty models converged successfully, therefore the Cox models are presented in the main tables.
Participants living in the lowest-SES areas tended to be older and a lower proportion was male, Caucasian, married or reported household incomes of at least $40 000 at baseline compared to those in the highest-SES areas (table 1). The median average household income for all block groups in the catchment area was $51 141 and the median proportion who completed college was 27.1%. By comparison, 51.9% of VITAL participants reported household incomes of less than $60 000 per year, and 42.3% completed college (data not shown).
Table 2 gives the mean, median and range of the measures included in the area-level SES index for all block groups in the 13 counties of the Western Washington SEER registry, and the block group-level measures for VITAL respondents. The overall catchment area and VITAL participants’ block groups both represented a wide range of SES; however, on average, VITAL respondents’ block groups had higher household incomes and home values, and a higher proportion of residents who completed high school and college, who were in professional and managerial occupations, and who lived in households that received net rental, interest or dividend income.
Table 3 gives total and site-specific incidence rates per 10 000 person-years and HRs and 95% CIs by quintile of area-level SES. Total cancer incidence was 144/10 000 person-years, which is somewhat higher than in all of the Western Washington SEER catchment area (116.7/10 000), driven largely by higher prostate cancer incidence in VITAL (79.7/10 000 vs 47.8/10 000).28 Total cancer incidence ranged from 135/10 000 in the highest-SES areas to 154.1 in the lowest-SES areas.
After controlling for demographics, living in the lowest-SES areas was marginally associated with higher total cancer incidence (HR: 1.08, 95% CI 0.99 to 1.17; Ptrend=0.067) and with higher risk of lung (HR: 2.21, 95% CI 1.69 to 2.90; Ptrend<0.001) and colorectal cancer (HR: 1.52, 95% CI 1.11 to 2.09; Ptrend=0.003; table 2, model 1). Prostate cancer risk was inversely associated with area-level SES (Ptrend=0.015). Area-level SES was not associated with incidence of breast cancer or of other cancers combined.
In models further adjusting for individual education and household income, the association between area-level SES and total cancer incidence attenuated (quintile 1 (Q1) vs quintile 5 (Q5) HR: 1.06, 95% CI 0.97 to 1.16; Ptrend=0.22) and was eliminated for area-level SES and prostate cancer (Ptrend=0.66) (table 2, Model 2). Living in the lowest-SES areas remained associated with higher lung cancer incidence (Q1 vs Q5 HR: 1.43, 95% CI 1.07 to 1.91) and marginally associated with higher colorectal cancer risk (Q1 vs Q5 HR: 1.35, 95% CI 0.97 to 1.88; Ptrend=0.062), particularly among men (Q1 vs Q5 HR: 1.53, 95% CI 0.99 to 2.38; Ptrend=0.031).
Table 4 gives cancer mortality rates, HRs and 95% CIs by quintile of area-level SES for total cancer mortality, and stratified by sex and by whether respondents were diagnosed with cancer before baseline. The overall cancer mortality rate was 48 deaths per 10 000 person-years and ranged from 33.6 in participants living in the highest-SES areas to 63.7 among those in the lowest-SES areas.
In models adjusted for demographics, living in lower-SES areas was associated with higher cancer mortality (Q1 vs Q5 HR: 1.68, 95% CI 1.47 to 1.93; Ptrend<0.001; table 4, model 1). The association between area-level SES and cancer mortality was somewhat weaker in respondents who were diagnosed with cancer before baseline (Q1 vs Q5 HR: 1.54, 95% CI 1.26 to 1.87; Ptrend<0.001) than among those not diagnosed before baseline (Q1 vs Q5 HR: 1.81, 95% CI 1.52 to 2.16; Ptrend<0.001). Controlling for individual SES substantially weakened these results; however, living in lower-SES areas remained associated with higher cancer mortality among all respondents (Q1 vs Q5 HR: 1.28, 95% CI 1.11 to 1.48; Ptrend<0.001) and particularly among those not diagnosed before (Q1 vs Q5 HR: 1.40, 95% CI 1.16 to 1.69; Ptrend<0.001). These associations were similar among men and women.
The purpose of this study was to estimate the association between area-level SES and total and site-specific cancer incidence and total cancer mortality, and to assess whether observed associations remain after controlling for individual SES. Area-level SES was inversely associated with lung and colorectal cancer incidence and total cancer mortality. Controlling for individual SES weakened these associations; however, area-level SES remained associated with lung cancer incidence and total cancer mortality, and marginally associated with colorectal cancer risk, suggesting that there are moderate-to-large associations between area-level SES and specific cancer outcomes, which are not completely explained by individual SES.
While measures of area-level SES should summarise information about socioeconomic conditions in a given area in a meaningful way and use data that can be compared between different locations and at different times,8 there are no established standards for measuring area-level SES, making it difficult to directly compare results between studies. However, previous studies have also used categorical area-level SES measures, allowing for comparisons of relative SES and cancer outcomes.
Several prior studies have included measures of individual SES in multivariate-adjusted models of area-level socioeconomic factors and cancer outcomes;1 ,2 ,5 ,6 ,9 ,10 ,20–22 however, very little prior work has directly compared these associations with and without control for individual socioeconomic factors,6 ,10 or presented results controlling for individual SES without simultaneously adding several additional risk factors.5 ,20 In a case–control study of area-level SES and prostate cancer among Caucasian and African-American men in South Carolina, Sanderson et al6 reported an OR of 0.52 (95% CI 0.34 to 0.80) associated with living in ZIP codes in the highest quartile of area-level SES relative to the lowest. Adding individual educational attainment resulted in wider CIs but had little impact on the effect estimates. These results differed from our findings of lower prostate cancer incidence in lower-SES areas before accounting for individual SES, but no association once individual education and income were included. These differences may be due to the higher proportion of African-American men included in the Sanderson et al study, and possibly to high levels of screening in VITAL.
A previous study of area-level SES and premature cancer mortality among individuals ages 25–64 in Australia reported an age-adjusted rate ratio (RR) of 1.69 (95% CI 1.54 to 1.84) comparing the most disadvantaged quintiles of Statistical Local Areas to the least in men, which attenuated to 1.48 (95% CI 1.35 to 1.63) after adding individual occupation, similar to our results.10 Associations between area deprivation and cancer mortality were much weaker among women (RR: 1.31, 95% CI 1.19 to 1.44) and did not change when including individual occupation.
In analyses adding parish-level unemployment to models already accounting for individual demographics and SES, proportion unemployed was inversely associated with lung cancer incidence in Denmark, similar to our results, and positively associated with prostate cancer incidence, unlike our findings of no association after including individual SES; however, results without controlling for individual SES were not presented.5
A case–control study in Wisconsin reported an OR of breast cancer of 1.23 (95% CI 1.09 to 1.39) for women in the highest quintile of census tract-level SES relative to those in the lowest quintile accounting for individual educational attainment.20 In contrast, we observed no association between area-level SES and breast cancer incidence when accounting for individual SES, consistent with other previous findings of no association between area-level SES and breast cancer risk after including individual education and other risk factors.21 ,22
Analyses in the Nurses’ Health Study found lower incidence of rectal cancer in women living in the highest-SES areas (using quintiles of the same index as in this study) relative to women in the lowest-SES areas when accounting for educational attainment and several other area-level characteristics and individual risk factors (relative risk: 0.64, 95% CI 0.44 to 0.93).2 This same study reported no association between area-level SES and colon cancer overall.2 We found no association between area-level SES and combined risk of colon and rectal cancer in women; however, we did not have sufficient numbers of cases to examine colon and rectal cancer separately.
Additional studies of area-level SES and cancer outcomes have included individual socioeconomic factors along with several modifiable risk factors (eg, diet, physical activity, smoking, obesity, alcohol use) that are likely on the causal pathway between area-level SES and cancer outcomes.1 ,9 Major et al9 reported higher cancer mortality rates in quintiles of the lowest- relative to the highest-SES census tracts in the NIH-AARP Diet and Health Study. This association attenuated substantially in models including individual education; however, these models also included demographics, family history, and several behavioural risk factors that could be on the causal pathway between area-level SES and cancer mortality, making the contribution of individual-level SES unclear. Similarly, Doubeni et al1 reported an inverse association between quintiles of area-level SES and colorectal cancer incidence that weakened when including individual education in addition to measures of diet, physical activity, body mass index, and smoking, which represent pathways through which area-level SES could impact colorectal cancer risk.4
Although previous work suggests that individual behaviours, such as smoking, alcohol consumption, diet and physical inactivity partially explain differences in cancer outcomes by area-level SES, the association between area-level SES and cancer mortality could also be due to higher incidence of more fatal types of cancer (eg, lung cancer, as found in this study); delay in seeking care for symptoms or less screening, both of which would lead to later stage at diagnosis; and/or less access to effective medical treatment or lower compliance with treatment, which could lead to poorer survival. Associations between area-level SES and cancer incidence and mortality that remain after controlling for individual SES (compositional factors) could indicate that there are also area-level (contextual) effects on these cancer outcomes, either directly or through influences on the factors noted above, suggesting that cancer prevention interventions may be important at the individual as well as area levels.
Limitations of this study should be noted. Although associations between area-level SES and cancer outcomes remained after controlling for two measures of individual SES, the remaining association could be at least partly due to residual confounding caused by measurement error in individual education and household income and by not including other measures of individual SES (eg, total assets; lifecourse socioeconomic factors) if they are more important. Our analysis of cancer mortality is limited because analyses of those diagnosed before baseline could be impacted by survival bias. Participants diagnosed before baseline had to survive long enough to be included in the study, and individuals from low-SES areas who are included here might not be representative of all individuals from lower-SES areas. There are differences in the types of cancer experienced in each group, with those diagnosed after baseline dying predominantly of rapidly-fatal cancers such as lung, pancreatic and haematological cancers and not breast, prostate or colorectal cancers. Additionally, models incorporating interaction terms between area-level SES and individual SES could add further insight into whether the relationship between area-level SES and cancer outcomes is consistent by different levels of individual SES.
VITAL recruited participants from only one region of the USA and a large majority of participants were Caucasian, which could limit its generalisability to other populations. Although baseline addresses were successfully geocoded for almost all VITAL respondents, misclassification of quintile of area-level SES index would occur to the extent that participants were placed in the wrong block group, and that particular block group was in a different quintile than the respondent's actual block group. Sensitivity analyses randomly reassigning quintiles of block group SES to 2% of VITAL participants resulted in only small alterations in the association between area-level SES and cancer incidence and mortality, suggesting that misclassification of area-level SES likely had only a small impact on our results. Area-level SES was assessed only at baseline and might not accurately reflect area-level SES at other aetiologically-relevant time points.
Strengths of this study include its prospective design, large sample size, and several years of follow-up, allowing for examination of several site-specific cancers as well as total cancer incidence and cancer mortality. Information collected from the detailed baseline questionnaires allowed us to control for demographic factors and to include two measures of individual-level SES. Linkage with SEER and the Washington State death file allowed for accurate and near-complete ascertainment of new cancer diagnoses and cancer deaths.
To the best of our knowledge this is the first study to systematically examine associations between area-level SES and total and site-specific cancer incidence and total cancer mortality with and without control for individual SES. Behaviours and other modifiable factors that affect cancer outcomes may be influenced by the socioeconomic characteristics of individuals as well as the socioeconomic and physical characteristics of the neighbourhoods in which they live. Future research should examine the causal pathways linking lower area-level SES to specific cancer outcomes to identify potential points of intervention to reduce these disparities.
What is already known on this subject
Living in lower-socioeconomic status (SES) areas has been associated with higher incidence of some cancers and with higher cancer mortality.
These associations vary by cancer site.
The extent to which observed associations between area-level SES and cancer outcomes are due to individual socioeconomic factors is unclear.
What this study adds
Living in low-socioeconomic status (SES) areas was associated with higher total, lung and colorectal cancer incidence, and higher total cancer mortality.
After accounting for individual education and household income, living in lower-SES areas remained associated with higher lung and colorectal cancer incidence, and higher total cancer mortality.
Associations between area-level SES, and cancer incidence and mortality, are partly explained by individual SES, but the places people live could also influence cancer outcomes, either directly or through other risk factors.
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Files in this Data Supplement:
- Data supplement 1 - Online supplement
Contributors TAH and EW contributed to the conceptualisation of the study, acquisition of data and the interpretation of the results. TAH analysed the data and wrote the manuscript. TAH, SAAB, LS and EW contributed to the analysis plan, provided critical review and contributed to revising and finalising the manuscript.
Funding This work was financially supported by the Biobehavioral Cancer Prevention and Control Training Program at the University of Washington funded by the National Cancer Institute R25CA92408 (to TAH) and by the National Cancer Institute and the National Institutes of Health Office of Dietary Supplements K05CA154337 (to EW).
Competing interests None.
Ethics approval Fred Hutchinson Cancer Research Center.
Provenance and peer review Not commissioned; externally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.