|Home | About | Journals | Submit | Contact Us | Français|
Although numerous studies have examined the association of area socioeconomic status (SES) and cancer screening after controlling for individual SES, findings have been inconsistent. A systematic review of existing studies is timely in order to identify conceptual and methodologic limitations and to provide a basis for future research directions and policy.
The objectives were to: 1) describe the study designs, constructs, methods, and measures; 2) describe the independent association of area SES and cancer screening; and 3) identify neglected areas of research.
We searched 6 electronic databases and manually searched cited and citing articles. Eligible studies were published before 2008 in peer-reviewed journals in English, represented primary data on individuals aged ≥18 years from developed countries, and measured the association of area and individual SES with breast, cervical, or colorectal cancer screening.
Of 19 eligible studies, most measured breast cancer screening. Studies varied widely in research design, definitions and measures of SES, cancer screening behaviors, and covariates. Eight employed multilevel logistic regression, the remainder analyzed data with standard single level logistic regression. The majority measured 1 or 2 indicators of area and individual SES; common indicators at both levels were poverty, income, and education. There was no consistent pattern in the association between area SES and cancer screening.
The gaps and conceptual and methodologic heterogeneity in the literature to date limit definitive conclusions about an underlying association between area SES and cancer screening. We identify five areas of research deserving greater attention in the literature.
Cancer is the second leading cause of death among U.S. adults, accounting for approximately one-fourth of all deaths in recent years (1). Of 1.5 million estimated new cases of cancer in 2009, 24% will be breast, colorectal, and cervical cancers (2), the three cancers for which broad consensus exists for population screening (3, 4). Yet, there remains a gap between actual and ideal rates of routine screening for these cancers in the U.S., particularly for colorectal cancer (CRCS) (5, 6). Low rates are especially troublesome for populations of lower socioeconomic status (SES) because they experience disproportionate cancer incidence and mortality (5, 7, 8).
Socioeconomic status, widely defined as the social and economic standing of an individual within a hierarchically-stratified society, has consistently been associated with cancer screening rates: low-income, less educated and working class populations are less likely to have obtained and maintained cancer screening than their counterparts (6, 8–11). Although widely referred to as “health disparities,” we embrace the principle of distributive justice (12) and its underlying tenet that disparities in health between the advantaged and disadvantaged are unnecessary, unjust, and avoidable (13), by referring to these SES differentials as social inequities in health. The consistent findings regarding cancer inequities have led to numerous programmatic responses, including educational interventions aiming to reduce SES-associated inequities in knowledge and attitudes and publicly-funded programs providing screening to low-income and uninsured individuals (14–17).
SES inequities in cancer incidence, mortality, and screening, have persisted despite decades of research and interventions targeting individuals (7, 18–20). Because cancer screening requires interaction with both healthcare services and the larger environment in which those services exist, those seeking to understand and positively influence these behaviors have increasingly turned to multilevel or contextual frameworks (21–24). These frameworks move beyond the individual and conceptualize health and health behaviors as a product of the dynamic interrelation of multiple levels of influence, including the individual, community, social, and structural. For example, multilevel or “ecosocial” (25) theories, such as neighborhood social disorganization (26) and social capital (27) and those models currently developing in cardiovascular and glycemic disease (28–31) hypothesize about the direct and indirect effects of social and contextual influences as measured by variables such as neighborhood disorder, social cohesiveness, residential mobility, area-level poverty, or the availability of parks and recreational facilities. In contrast, individual-level frameworks conceptualize health behaviors and medical outcomes as products of such individual-level attributes as knowledge, attitudes, or genetics. By facilitating a clearer understanding of the pathways and mechanisms linking geographic areas and health, multilevel frameworks offer the promise of new and innovative interventions and policies to increase cancer screening and reduce cancer-related inequities.
A growing number of studies have used multilevel frameworks to examine the association of area-level SES—the economic, educational, occupational, or class status of a particular area— and cancer screening. Several studies have demonstrated relations between areas characterized as low in SES and a greater likelihood of late- or no- screening, even after controlling for the contribution of individual SES (23, 32–38). Not all of the measured area-based socioeconomic measures (ABSMs) have been significantly associated with screening (23, 33); in other studies, significant associations have become non-significant following adjustment for individual SES (39–41). Moreover, at least one study found no variation in cancer screening at the area-level (42).
Conceptual and methodologic differences in this literature may be responsible for the observed variation in findings. Moreover, these differences may limit comparability across studies, hinder interpretation of observed associations between area-level SES exposures and cancer outcomes, and constrain our ability to develop and test hypotheses about potential mechanisms linking neighborhoods, screening behaviors, and cancer outcomes.
To provide coherence to this emerging and heterogeneous research area, we conducted a systematic review to describe studies of area SES and individual cancer screening behaviors. Specific aims were to: 1) describe the study designs, constructs, methods, and measures; 2) describe the independent association of area socioeconomic status (SES) and cancer screening over and above individual SES; and 3) identify neglected areas of research. While common methodologic practice in the systematic review literature is to pool results from homogenous groups of high quality studies in order to synthesize the literature and draw conclusions regarding the strength of the association of interest using a narrative best evidence synthesis framework (43) or quantitative meta analytic techniques (44) the high degree of observed heterogeneity in these studies precluded our ability to summarize and draw definitive conclusions about the underlying association of interest. We do, however, outline a summary of the tested associations and their outcomes in the literature to date in the results section and leverage both the identified sources of heterogeneity and neglected areas of research as a basis for our discussion, in which we identify the key priority areas needing refinement in the multilevel cancer screening literature.
Studies were included if they: 1) were written in English; 2) were published or in-press in a peer-reviewed journal; 3) reported data from a primary study (not a review or editorial); 4) were conducted in developed countries; 5) included participants aged 18 years and older; 6) measured area SES with one or more ABSMs as an independent variable and provided one or more measures of effect; 7) measured cancer screening uptake (cervical, breast, or colorectal) at the individual-level as a dependent variable; and 8) controlled for one or more measures of individual SES in the analysis of ABSM effects. Eligibility criteria were applied hierarchically in the numeric order above. Studies using either multilevel analyses (e.g., hierarchical logistic regression methods that control for the clustering of individuals within geographic areas) or single level analytic methods (e.g., standard logistic regression methods that do not account for the nested structure of the data) were eligible for inclusion.
ABSMs were broadly defined to include compositional (measures of central tendency aggregated from multiple individual-level observations) or contextual (area-level observations) measures of SES within a defined geographic area ranging in size from as large as a state or province to as small as a block group. Individual SES measures were defined as individually-reported characteristics such as social class, occupation, Medicaid status, education, or individual or household income. Cancer screening was defined as tests that are recommended for population-wide screening for breast, cervical, or colorectal cancer (3, 4). Developed countries were those classified as having “high human development” on the United Nations Human Development Index (45).
We searched six electronic databases that index behavioral and health journals to identify peer-reviewed articles meeting our inclusion criteria published through the year 2007. The databases, the platform, and the dates they began coverage include: Medline (via OVID) 1950, CINAHL (via OVID) 1982, Academic Search Premier (via EBSCO) 1990, Sociological Abstracts (via CSA) 1963, Social Services Abstracts (via CSA) 1980, and PsycINFO (via OVID) 1887.
An informal and brief review of the literature was used to identify preliminary keywords. The keywords fell into 3 broad categories (Appendix A): area (e.g., census tract, metropolitan area), socioeconomic factor (e.g., poverty, deprivation), and cancer screening test (e.g., Pap smear, mammography) and were adapted for each database’s unique thesaurus of subject headings with the assistance of a public health research librarian. The search strategies were pretested and refined in an iterative process in which citations were screened for relevance and for additional keywords and exclusion criteria. The search was modified as necessary to ensure that it captured the sample of eligible articles obtained by the first author through initial literature searches and previous research. The final search strategies include keywords and subject headings specific to each of the electronic databases (Appendix B).
For each eligible article, we searched the article bibliography, entered the citation into the ISI Web of Science database, and contacted first authors to identify additional eligible cited, citing, or new articles. Finally, we searched Proquest Digital Dissertation database and contacted all authors of potentially eligible dissertations to determine whether the dissertation had been recently accepted for publication in a peer reviewed journal.
We employed a two-step process for screening studies. First, titles and abstracts of all articles were screened by the first author to determine eligibility. Next, for all abstracts meriting further review and for those abstracts in which eligibility was unclear, full texts were reviewed for eligibility. A quality control check was conducted using an independent reviewer (M.S.) who reviewed titles and abstracts of a random sample of 5% (n=38) of the excluded and 50% (n=9) of the included articles. He and the first author disagreed on 2 excluded and 2 included articles. Upon reviewing the full text of these 4 articles, however, the independent reviewer agreed with the first author’s decisions.
In data extraction, all eligible studies were read and coded by two reviewers: the first author (SLP) and one member of the review panel (the co-authors and an MPH-trained research assistant). A web-based systematic review software program, SRS 4.0 (TrialStat, Ottawa, Canada), was used for data extraction. All codes were checked for interrater reliability by examining the percent agreement (0.92) and kappa statistic (0.82). Discrepancies were resolved by discussion and consensus between the coders.
Consistent with our aims and an earlier review of multilevel studies (46), we identified study characteristics and findings for extraction. For aim 1, we extracted data on the study sample, design, data sources, area-level, method of analysis, and measurement of individual and area SES, cancer screening outcome(s), and covariates. For aim 2, we coded information on the association of ABSMs and cancer screening by extracting data from a maximum of 2 analytic models per dependent variable, including the effect sizes and confidence intervals for the unadjusted (or least adjusted) and the final fully adjusted models, as available. Finally, for aim 3, we evaluate the existing literature and identify neglected areas of research to provide direction and guidance to researchers designing the next generation of multilevel cancer prevention research.
The database search identified 771 potentially relevant unique citations, of which 12 (1.6%) met our inclusion criteria. The additional steps in the search led to the identification of another 7 eligible articles, for a total of 19 (Figure 1). The majority (n=13) of the studies measured breast cancer screening, including mammography, clinical breast exam, and breast self-exam (21, 32–36, 38–40, 47–50) (Table 1). Publication dates ranged from 1998–2007, and the majority of studies were conducted in the U.S. (21, 23, 33–35, 37–40, 42, 47, 48, 50–52). Fifteen studies employed cross-sectional analyses (21, 23, 32, 33, 36–38, 40–42, 47, 49–52). The majority (n=11) analyzed data at a single level using the standard analytic technique of logistic regression (21, 34–38, 41, 47, 48, 50, 51); the remaining 8 articles employed 2-level multilevel logistic regression models (23, 32, 33, 39, 40, 42, 49, 52), with one of these studies also estimating a 3-level model (52) (Table 1).
Eight studies linked self-report individual-level data obtained from surveys to separate administrative sources of area-level data such as the U.S. census (21, 36, 39, 41, 42, 47, 51, 52). Three studies used the same data source for both individual- and area-level data, using aggregated individual-level data to represent area-level exposure (32, 38, 40). Four studies provided details about the geocoding procedures used to link individual- and area- level data such as rates of address matching (35, 39, 48, 52). Eleven included national samples (21, 23, 36, 38–42, 47, 51, 52). In eight of twelve studies reporting racial/ethnic composition, populations consisted of 75% or more non-Hispanic whites (21, 23, 33–35, 40, 42, 50). Although a few studies reported including large (≥25%) (48) or entirely African American populations (39, 52), no studies reported including significant populations (≥25%) of Hispanics.
While inclusion criteria varied, all studies had some age requirement and several (n=4) restricted the sample to individuals with no personal history of cancer (21, 42, 48, 52) or no current cancer or symptomatic screening tests (34, 39). A few studies (n=4) also applied inclusion/exclusion criteria to the area-levels of interest (21, 23, 40, 51) for example, limiting study inclusion to residents living in areas with a population ≥1.5 million.
The dates of data collection for individual- and area-level data were not consistently or clearly reported in five studies (21, 33, 37, 42, 47). Exposure (area-level) data were measured between 1 and 8 years prior to outcomes in 9 studies (23, 33, 34, 36, 41, 48, 50–52) and at the same or approximately the same time as individual outcomes in 6 studies (32, 35, 38–40, 49).
All studies defined geographic “areas” using some form of country-specific administrative boundaries (e.g., metropolitan statistical areas (MSAs), counties, census tracts, collection districts) or best approximations of administrative areas (38) rather than resident-defined “neighborhoods” or places delineated by other spatial criteria, such as the radius around a respondent’s home. The primary source of area-level data in U.S. studies was the U.S. Census (n=8) (33–35, 39, 48, 50–52) and most U.S. studies analyzed data aggregated within relatively large areas, including MSAs (n=2) (21, 40), counties (n=5) (23, 33, 42, 47, 51), and zip codes (n=3) (34, 35, 50). Fewer studies analyzed data at the census tract level (n=2) (48, 52) (population range: 1500–8000) or block groups level (n=1) (39) (population range: 600–3000).
We identified and coded multiple ABSMs used in these 19 studies, described in detail in Table 1. Most studies used just one (n=12) (21, 32, 34–37, 39–41, 47, 50, 52) or two measures (n=3) (23, 49, 51). The majority of studies (n=12 of 18 reporting) modeled ABSMs as categorical variables (34–39, 41, 47, 48, 50–52). Most studies measured at least one indicator of percent living in poverty (n=10) (23, 33, 37, 38, 40, 42, 47, 48, 51, 52), area income (n=6) (32, 34, 35, 38, 42, 48), or area education (n=7) (21, 23, 38, 48–51). Fewer articles measured unemployment or occupation (n=3) (38, 48, 49), SES indices (n=4) (36, 39, 41, 48), or other SES indicators (n=2) (33, 48). Poverty was defined following U.S. government standards in which total family earnings are compared to the U.S. poverty thresholds, standards set for families of certain ages and sizes. Families under the threshold are considered to live in poverty (53). Most measures of income captured median household, family, or per capita income, representing measures of area central tendency rather than measures of income distribution. Whereas income was measured as an absolute phenomenon, education was typically measured as a relative phenomenon, with the percent of individuals above or below a certain level (e.g., percent of population with high school education or above). A few studies (n=4) measured indices comprised of multiple SES indicators including education, income, and unemployment (36, 39, 41, 48). No studies measured socioeconomic inequities—unequal distributions of economic assets, income, or education—using measures such as the Gini coefficient.
Individual SES was controlled for using 1 to 4 measures. Table 2 displays the individual SES measures used and the number of categories for each measure. The most commonly used indicators were education (n=15) (21, 33, 34, 36–42, 47–49, 51, 52) and income (n=9) (21, 32, 37–40, 42, 48, 51); three studies calculated federal poverty levels using household income and the standard methods described above (33, 38, 47). Income was measured with annual family or household income and only one study controlled for family size or the number of people supported by that income (51). Education was typically defined as the highest level of education obtained. Three studies of Medicare beneficiaries used dual enrollment in Medicare and Medicaid as a proxy measure of individual low-income status, presumably due to a lack of other available SES indicators (23, 35, 50). Other individual SES measures included employment status (e.g., unemployed vs. employed) and occupation (e.g., professional vs. nonprofessional).
Individual-level covariates included sociodemographic characteristics such as age, race/ethnicity, gender, and health insurance status. Behavioral or psychosocial covariates were rarely measured. Nine studies measured area-level covariates (21, 23, 32, 33, 35, 47, 48, 50, 52) such as racial/ethnic composition, area medical services, or urbanity. Area racial/ethnic composition was conceptualized as the percent of population that was non-white, non-Hispanic white, Mexican, Asian, or black. Measures of area-level medical characteristics included HMO market share, the number of generalists or specialty physicians in an area, or medically-underserved areas.
Studies measured multiple dichotomous dependent variables encompassing 7 types of cancer screening (mammography, clinical breast exam, breast self exam, fecal occult blood test, sigmoidoscopy, colonoscopy, and Pap smear) for breast, cervical, and colorectal cancer. Cancer screening uptake was measured by self-report in most studies or by clinic, administrative, or claims data in 5 studies (23, 34, 35, 48, 50); no studies used both methods. One study reported excluding testing done for diagnostic reasons or symptoms (23). The majority of dependent variables in the cross-sectional studies measured recent or adherent screening (within previous 1–2 years for breast cancer screening, 1–3 years for cervical cancer screening and 1–10 years for colorectal cancer screening) (21, 23, 32, 33, 36, 37, 41, 42, 47, 50–52). Three studies measured lifetime screening only (38, 40, 49). The 3 prospective studies (34, 39, 48) and 1 retrospective study (a study measuring screening retrospectively over time using claims data) (35), all of which measured breast cancer screening, measured repeat screening.
In all, 132 unadjusted and adjusted associations between ABSMs and cancer screening were tested (Table 2). Seventy-one (53.8%) associations were statistically significant and positive, indicating that as SES increased (i.e., increasing income and education, decreasing poverty and unemployment), the odds of cancer screening increased. Only 6 (4.5%) were negative associations and 55 (41.7%) were not statistically significant. Of 54 tested associations from models that did not control for individual SES, 38 (70.4%) were positive, 2 (3.7%) were negative, and 14 (25.9%) were not statistically significant. After controlling for individual level SES and other covariates in fully adjusted models, 32 (41.0%) of 78 tested associations were positive, 4 (5.1%) were negative, and 42 (53.9%) were not statistically significant. Fewer associations were statistically significant and positive in multilevel models (n=17, 43.6%) versus single level models that did not account for the clustering of individuals in areas (n=53, 57.0%).
Notably, while there are many positive and nonsignificant results, there are far fewer negative associations, with the majority of those attributable to one study. In this Japanese study, the authors suggest that the unexpected direction of effect may reflect a lower likelihood of screening in urban areas resulting from the use of an imprecise indicator of urbanity and/or the very strong correlation between urbanization and higher regional indicators of income and education in Japan (32).
In all, 13 studies measured breast cancer screening; 4 employed prospective or retrospective study designs (34, 35, 39, 48) and 5 provided results of multilevel analyses (32, 33, 39, 40, 49). Of all studies, 2 demonstrated only nonsignificant results (21, 47), 1 demonstrated only negative associations (32), 2 demonstrated significant interactive effects (49, 50), and the remaining 8 studies (including the 4 prospective or retrospective studies) demonstrated at least one significant positive association, although in a prospective study, one of these was no longer significant after adjustment for household income (39) and another prospective study demonstrated both positive and negative associations between categories of zip code median income and mammography adherence (34).
Of the 2 studies demonstrating interactive effects, a Canadian study demonstrated a significant cross level interaction between regional education and social involvement, demonstrating that the influence of increased participation in volunteer activities and religious services modified the negative influence of low regional education level on mammography uptake (6). In a study of U.S. Medicare beneficiaries, compared to those living in zip codes in the highest quintile of college education rates, Medicare-only women living in the lowest quintiles were less likely (OR: 0.73; 95% CI: 0.72–0.74) and Medicare-Medicaid women were equally as likely (OR: 0.98; 95% CI: 0.94–1.02) to have had a recent mammography (41).
Several breast cancer screening studies measured multiple associations, encompassing multiple ABSMs and/or dependant variables (33, 38, 41, 48). One study measured a total of 40 associations including both unadjusted and adjusted associations between a single dependent variable and 10 ABSMs separately for African American and white women. In this single level prospective study, different ABSMs predicted nonadherence to mammography for blacks and whites. In the fully adjusted models of 10 ABSMs, 5 demonstrated statistically significant positive associations with mammography: 3 among blacks (% working class, % adults without high school education, and the SES Index) and 2 (% living in crowded conditions, % of homes valued ≥$300,000) among whites (48).
Of the 3 cancer screening modalities included in this review, breast cancer screening is the best developed and most frequently studied. The findings of these studies, while mixed, demonstrate predominantly positive associations. Several studies demonstrated that the associations between area SES and breast cancer screening may differ by individual or area characteristics or by ABSM (33, 38, 48–50) and several studies demonstrated attenuation of effects after control for covariates (38, 39, 48). Overall, these results indicate that associations between area SES and breast cancer screening may be more complex than previously acknowledged and suggest a need for further research in ABSM measurement and both moderating and mediating pathways in the breast cancer literature.
Eight cross-sectional studies examined cervical cancer screening (21, 32, 38, 40, 41, 47, 51, 52). These studies typically controlled for few individual or area covariates, 5 measured only a single ABSM (21, 32, 40, 41, 47, 52), and only 3 employed multilevel analyses (40, 52, 54). Two studies examined only lifetime use of a Pap smear (38, 40) and the others examined Pap screening in the last 1–3 years. One of the U.S. studies was conducted at the census tract level (52); no studies were conducted at the block group level.
In the 4 studies presenting unadjusted associations, 1 found only negative associations (32) and 3 found at least one significant positive association between lower area SES and lifetime receipt of Pap screening (38, 40, 41). In models that controlled for individual SES, only 2 of 8 studies demonstrated at least one significant positive association between low area SES and Pap screening and 1 study demonstrated 1 negative and 1 significant interactive effect. After controlling for all individual covariates and state level poverty, a multilevel study of black women found that compared to those living in census tracts with <5% poverty, women living in high poverty areas (≥20% poverty) were 1.2 times more likely (95% CI: 1.1–1.4) to have not had a Pap smear in the 2 years (52). The other study with positive adjusted associations found larger effects between area median income, percent living in poverty, and lifetime receipt of Pap smear but controlled for fewer covariates (38). A single level study found a negative association between area poverty and Pap smear in the last 3 years, a non-significant interaction between individual income and county-level poverty and a significant interaction between individual and area-level education, with less educated women living in better educated counties less likely to undergo Pap screening than the same women living in less educated counties (51). One additional study reported the results of testing for interactions and found a non-significant interaction of census tract poverty rate and individual education(52).
Overall, the findings for cervical cancer screening were mixed and demonstrated positive, negative, nonsignificant, and interactive associations. Compared to the breast cancer literature, the cervical cancer literature is sparse and includes few studies conducted at smaller, more homogenous geographic levels and few multilevel studies.
Five cross-sectional studies measured CRCS (23, 32, 37, 40, 42), 4 of which employed multilevel analyses (23, 32, 40, 42). All but one of the CRCS studies (32) were conducted in the U.S., none examined census tracts or block groups, 1 measured lifetime use of screening (40), 2 measured recent or adherent screening within the last 1–10 years, with the time interval contingent on test modality (37, 42) and 2 measured testing within the last year regardless of test modality (23, 32).
In all, 3 studies demonstrated at least 1 significant positive association between area SES and CRCS, 1 study demonstrated a significant negative association, and 1 study found only nonsignificant associations. In the 2 studies that controlled for the most comprehensive sets of individual- and county-level covariates such as use of and local availability of healthcare services (23, 42), 1 study tested 8 associations and found 2 positive associations, with a higher county-level percent of adults with high school diplomas associated with greater likelihood of FOBT (OR: 1.06, 95% CI: 1.00–1.12) and SIG (OR: 1.14 95% CI: 1.08–1.20) in the last year. In the same models, poverty was not significant; and neither ABSM was associated with colonoscopy alone or a combined variable of either sigmoidoscopy or colonoscopy (23). The other study did not present data but reported no significant random variation at the county level and reported that none of the county-level variables (percentage of population in poverty, median family income, and per capital income) were associated with any type of CRCS adherence in the multivariate models (42). Of the remaining 3 CRCS studies, the Japanese study reported a negative adjusted association (32) and the other 2 U.S. studies found that all 6 of the tested associations between MSA- and neighborhood-level poverty and lifetime (40) and adherent screening (37) were positive.
Overall, with only 5 studies, the CRCS literature is the most limited domain of the 3 cancer screening modalities reviewed here. The findings are mixed and while the majority of studies use multilevel analyses, no studies measured SES at the census tract or block group level, and only 2 studies measured multiple ABSMs.
Overall, we found a high degree of heterogeneity in study design, statistical methods, and definitions and measures of area and individual SES, screening tests, and covariates. Given the limited sets of conceptually and methodologically similar studies in which to compare and contrast findings, and a lack of clear patterns in the associations by outcomes, study design, and methods, we were unable to provide a definitive evidence synthesis or draw conclusions about the strength of the evidence either for or against an area SES-cancer screening association over and above individual SES.
The broader area and health research literature has seen significant debate and discussion regarding current methodologic practices (55–66) and previous reviews and critiques have outlined multiple weaknesses in the literature (46, 67–70). Building on these critiques and incorporating the identified sources of heterogeneity and weaknesses in the studies reviewed here, we identify 5 key priority areas of research and set 6 immediate research targets for the multilevel cancer screening literature.
This literature is currently dominated by either atheoretical studies or individual-level behavior change models and is limited by a lack of understanding regarding the mechanisms linking areas and behavior. Only a handful of studies cited theory or conceptual models as justification for their selection of variables or analytic methods (33, 34, 40, 47, 48). Cited models include the Health Belief Model and Andersen’s Behavioral Model of Health Services Use. Many individual-level psychosocial and behavioral correlates and predictors of cancer screening behavior have been identified (71), including receiving a screening recommendation from a doctor and perceived risk of cancer; however, these variables were rarely included in the reviewed studies. Thus, while we have some understanding regarding the mechanisms linking individual SES and health behavior and outcomes (72), we know little about the pathways through which area SES influences health and screening behaviors. Several studies discussed possible mechanisms linking area characteristics and cancer screening by suggesting that access to material resources such as medical facilities was a primary driver of observed associations. Other studies suggested social influences and resources as possible mechanisms linking area characteristics to individual behavior. Suggested social mechanisms included social support, social cohesion, collective efficacy, social capital, social norms, values and attitudes about healthcare, social and economic pressures shaping the quality of healthcare options, and overall variability of opportunities for choices about healthcare behaviors in different areas. Many of these mechanisms, well-established in other fields, deserve hypothesis development and testing in the cancer screening literature. While some of the reviewed studies suggested that area characteristics may shape individual behavior by influencing individual income, education, and employment opportunities, few explored cross-level ABSM interaction terms and no studies tested for mediation.
Although several conceptual models have been developed, no comprehensive and testable theory is available that offers an adequate explanation of the mechanisms linking area-level factors to a diversity of health outcomes (69, 73, 74). The challenge for the cancer screening literature is to begin to integrate our established individual-level models and constructs with perspectives capturing higher-level structural inequities. Similar efforts at conceptual integration are already underway in other fields where researchers are mapping the links between areas and cardiovascular and glycemic disease (28–31). It is important for all prevention sciences to move beyond models identifying individual-level predictors of health outcomes to modeling the pathways linking areas and individual health behaviors.
Acknowledging the influence of different intra- and inter-national cultural contexts, rather than assuming homogeneity of area-level effects across populations and societies, will be of critical importance. In this review, for example, unexpected negative associations were seen in the one reviewed Japanese study (32), possibly the result of a different cultural context. And a recent cancer screening study demonstrated three-way interactions between area-level SES, gender, and race (75), suggesting differential effects of area-level factors by population subgroup.
The current lack of cross-level conceptual integration limits our ability to develop and test creative and appropriate hypotheses regarding mediating and moderating pathways. By incorporating hypothesized multilevel pathways and mechanisms, newer models will provide a useful guide for researchers, policy-makers and health promoters seeking to understand and influence the multilevel context of cancer screening in order to prevent cancer-related inequities.
Many of the studies in this review demonstrated some significant associations between ABSMs and cancer screening; however, most were cross-sectional and thus unable to establish causality. While cancer screening researchers have been encouraged to use consistent measures of consecutive or repeat on-schedule screening (76–79) due to the high prevalence of lifetime screening, particularly for breast and cervical cancer, several studies measured only lifetime screening, further obscuring any potential temporal associations. Even studies with retrospective and prospective designs did not fully capture the temporal associations between areas and individuals. For example, while a few studies may have carefully and prospectively measured individual adherence to screening, the exposure of interest, area SES, may have been measured several years earlier or even concurrently with individual-level covariates and outcomes.
Although the majority of cross-sectional multilevel studies, including most of those in this review, do not include individual and area data drawn from the same time period, partly due to feasibility issues such as the infrequent collection of area-level data (55), they still treat the data as cross-sectional. By ignoring the need for a time-lag for many exposures and the possibility of change at the area-level, such as gentrification, and at the individual-level, with individuals migrating in and out of areas over time, these studies miss important opportunities to model potential causal relations between areas and health. Collecting exposure and outcome data at the same time implies a zero time lag between area exposures and individual outcomes, an implausible scenario unless the area exposure is assumed to be stable over time (80). Three studies collected exposure and outcome data from the same time periods by using the same data source for both levels, with areas comprised solely of aggregated responses to the individual survey. In addition to assuming a zero time lag, this approach can result in same source bias—a spurious association related to the use of the same individuals’ self report assessments of both exposure and outcome (81). Moreover, unless compositional effects tests are conducted, it is not known whether modeling an aggregated variable at a higher level of analysis provides increased explanatory power as compared to modeling the original variable at the individual-level.
The preponderance of cross-sectional designs, lack of temporally matched data, and lack of appropriately time-lagged study designs in this literature clouds an already unclear etiologic picture between area SES and cancer screening. At a minimum, assumptions of area stability or change over time should be theoretically justified; for example, while immediate changes in transportation infrastructure are unlikely, very rapid socio-cultural changes may result from area gentrification or area decline. Going further, these assumptions may be tested by examining the stability of population composition over time. Future research should employ prospective designs that incorporate aspects of lifecourse and choice theories and should control for the effect of individuals moving into and out of geographic areas.
Studies primarily relied on larger area units defined for political or administrative purposes. Decisions about the most appropriate geographic levels for analysis are often driven by considerations of data availability, sample size, and area homogeneity, with larger areas such as zip codes and counties considered more heterogeneous than smaller areas such as census tract and block groups. Extreme within-cluster heterogeneity may limit the utility of central tendency measures such as median household income while total homogeneity within an area may preclude studies of area effects (46). Although debate exists about appropriate geographic levels of analysis (61, 63, 66), previous U.S. research has indicated that smaller units such as census tracts or block groups model socioeconomic gradients in some health outcomes more consistently than larger geographic units (82–85). Research in Australia has also supported the use of smaller geographic units known as collector’s districts rather than larger postcodes (86). Studies using smaller, more homogeneous geographic units such as census tracts and block groups are underrepresented in the U.S. literature, including the cancer screening literature reviewed here, and deserve more attention. Future research should explore the validity and reliability of various ABSMs across geographic units in studies controlling for individual SES or in studies examining cancer prevention behavior outcomes. Cancer prevention researchers should also begin to incorporate approaches such as conditional autoregressive models (87) and hierarchical geostatistical models (88, 89) as methods to help guard against the modifiable areal unit problem, in which aggregation of continuous spatial phenomena to administrative areas of arbitrary sizes and shapes may introduce bias.
Ideally, researchers should specify the hypothesized mechanisms underlying the association of interest and select areas that best capture those pathways of influence. For example, studies have shown that the effects of socioeconomic inequity on health using measures such as the Gini coefficient are best captured with the use of larger areas encompassing larger populations and greater heterogeneity (90, 91). If area effects are hypothesized to influence outcomes via characteristics of inter-personal mechanisms such as social support, capital, or norms, community-defined and natural neighborhood boundaries or administrative areas of smaller sizes may be the most appropriate choice of area-level. Alternatively, because social service, housing, and medical services are typically delivered at county or other administratively-defined levels, researchers hypothesizing about these mechanisms should consider using the relevant administrative boundaries.
There are multiple weaknesses in the measurement of area-level SES in this literature including the use of a limited set of ABSMs, potential residual confounding from individual SES, the exclusive use of compositional ABSMs, and the unknown effects of categorizing ABSMs.
While multiple ABSMs were used in these studies, the majority were indicators of poverty, income, and education; fewer measures captured area distributions of employment, occupation, resources, or wealth. A much wider array of ABSMs and other area-level indicators have been employed in sociology and in the wider public health literature (46, 70, 92). Some examples of these variables include: availability of health services, measures of medical underservice, distance to medical facilities, neighborhood social and physical disorder, residential mobility, wealth as measured by assets or non-salary income, social class, community socialization and involvement, social trust and cohesion, norms, collective efficacy, compositional demographics, family structure, and subjective and perceived area SES (46, 70, 92–94).
Several studies in this review controlled for only a single individual SES variable and many variables were measured with only 2 or 3 categories. A few used a measure of dual enrollment in Medicare and Medicaid as the only individual SES measure. While this measure is thought to “reflect on the individual’s low income status” (23), not all eligible low-income individuals are Medicaid beneficiaries and dually-eligible individuals may differ from Medicare-only enrollees; for example, they are more likely to be institutionalized and face greater health concerns (95). With limited or inaccurate control of individual-level SES, ABSMs may be more likely to capture the unmeasured aspects of individual SES. However, if individual variables are mediators on the causal pathway between area SES and individual outcomes, controlling for multiple indicators of individual SES may result in an underestimation of the true independent effects of area-level characteristics (57).
All ABSMs in these studies were compositional variables. Such aggregated variables are assumed to represent contextual characteristics of neighborhoods. Contextual variables measured at the area-level, such as observation of the number of broken windows in a neighborhood, have been recommended by some to better capture area characteristics (96) and others have argued for triangulation with the use of multiple methods of ABSM ascertainment (81, 97). However, the lack of data about the reliability and validity of various area characteristics combined with a lack of standardized national data of contextual area characteristics has limited the applicability of such measures for large-scale U.S. studies. Ideally, the decision between compositional and contextual variables should be driven by theory. A classic example of a theory-driven compositional variable is social norms, demonstrated in statistical analyses to have predictive power beyond that explained by individually reported norms alone (98).
SES is often conceptualized as a relative phenomena (99, 100) and categorizing either area- or individual-level SES may obscure the full gradient of SES inequities (101). However, the majority of studies in this review employed categorical ABSMs and in many studies, only a few of the measured categories were significantly associated with the outcome. Whether these findings are indicative of non-linear associations, threshold effects, or are artifacts of sample size in each of the categories or the size and homogeneity of each category, is unclear. To begin to address these unanswered questions, researchers should be clear about their theoretical or methodologic reasons for employing categorized or continuous ABSMs and make explicit their reasons for their decisions.
The limitations of the current practices involving ABSMs combined with a lack of reporting about the psychometric and distributional properties of these variables hamper definitive conclusions about the validity of these measures across or within cancer screening behaviors. Future cancer prevention research should explore a broader range of ABSMs than previously explored, should acknowledge the well-known drawbacks resulting from the use of SES indices (46, 83–85), should begin introducing contextual variables, and should explore and report the effects of categorization. In particular, smaller, in-depth pilot and qualitative primary data collection studies are needed in the cancer prevention literature to introduce and explore subjective measures of area context as perceived by neighborhood residents as well as systematic social observations made by outsiders (non-residents) (81, 93, 94, 102). And while this first generation of research has taken advantage of existing available data, future decisions about ABSMs should be guided by theory when possible.
We restricted our review to studies that are inherently hierarchical, with individuals living nested within areas and studies measuring both individual- and area-level SES. However, over half of the studies used standard single level logistic regression analyses that did not account for this clustering. It should be noted that one study presented single level analyses after testing for and finding no significant random effects at area-level (48), an intuitively appropriate but debatable strategy given the power and estimation difficulties related to variance component tests in multilevel logistic models (103). Failure to test for multilevel nesting and subsequently account for the hierarchical nature of data using multilevel analyses leads to negatively biased standard errors potentially resulting in spuriously statistically significant results (103). Multilevel analyses, on the other hand, account for clustering (non-independence) of individuals within areas and allow for the simultaneous modeling of the association of group and individual characteristics and any interaction between the two on outcomes.
Multilevel analyses have their own methodologic challenges previously described in earlier methodologic papers (57, 59, 65, 80, 104, 105). One of the critical challenges in the multilevel literature that was poorly addressed in these studies is the decomposition of variance in the outcome into within and between area components. The relative influence and importance of area vs. individual attributes on individual outcomes is an unresolved debate. We cannot shed much additional light on this topic, in part because most studies reviewed here based their claims about the relative importance of multiple levels on an examination of effect sizes and associated p-values of area- and individual- fixed effects, an inappropriate strategy. An alternative method for drawing conclusions about the relative importance of multiple levels of influence is the intraclass correlation coefficient (ICC), a measure expressing the proportion of the totalvariance in the individual outcome resulting from the influence of the area-level. However, while widely used and appropriate for use in multilevel linear models, debate exists about the appropriateness of the ICC in multilevel logistic regression models. Few studies in this review (32, 40, 48, 52) presented ICCs or other partitioned variance estimates. Alternative measures such as median and interval odds ratios that capture area variance have been suggested (64, 106–108) and while appropriate for dichotomous outcomes, are not yet widely used and should be adopted in this literature.
Cancer screening researchers should take advantage of recent technological innovations in Bayesian and spatial analytic methods (87, 89, 109, 110) that can incorporate temporal and spatial dimensions such as spatial adjacency and decay functions and dynamic agent-based models, computer representations of the multiple feedback loops between people and places that lead to spatial patterning of individual-level outcomes (111). Further development and use of multilevel structural equation modeling, methods allowing for more sophisticated testing of moderating and mediating pathways, is also needed.
We have offered multiple suggestions for improvement within our 5 identified priority areas of research and while all are important, we acknowledge that to some extent, our suggestions represent an ambitious wish-list of ideals. Given the emerging nature of this research, the limitations of existing area-level data, and the developmental stage of our more advanced analytic methods, we acknowledge the importance of identifying realistic and immediate next steps of this research. That said, we recommend the following 6 more immediate research goals: 1) Infuse more theory into the literature by applying existing theories and developing and testing new conceptual models. 2) Conduct prospective studies and embrace natural experiments when possible. 3) Move beyond poverty by comparing contrasting multiple ABSMs within studies. 4) Conduct qualitative and pilot studies to capture unique social and physical characteristics at the area-level and to explore areal units that go beyond aggregated administrative data and administrative boundaries. 5) Accept complexity by testing for cross-level interactions and population subgroup effects. And lastly, 6) Continue to develop and implement alternative analytic technologies and when using multilevel logistic models, expand reporting of model statistics, and when possible, measures of area-level variance and median and interval odds ratios.
Among the 3 cancer screening behaviors in this review, the breast cancer screening literature is the most extensive and the only literature that has employed prospective designs. Greater attention to cervical and colorectal cancer screening and prospective designs is needed, particularly among African American and Hispanic populations that have been largely overlooked in this literature.
Limitations of this review include that it was limited to published, peer-reviewed, English-language articles and to studies measuring at least one indicator of area SES. We may have excluded studies measuring other important area-level factors such as medical underservice or area racial/ethnic composition. Despite these limitations, this is the first attempt to systematically review the literature in this emergent area of cancer prevention research.
Early efforts to monitor and promote cancer screening focused on individual determinants. In recent years, the cancer screening literature has witnessed an increasing number of calls to increase and improve research into the multilevel context in which health behavior occurs (78, 112, 113). Although conceptual and methodologic heterogeneity and weaknesses limited our ability to draw definitive conclusions about the underlying relations between area SES and cancer screening, we believe that awareness of these limitations will lead to better decisions regarding conceptualization, measurement, and analysis of area SES factors associated with cancer screening. Addressing these issues will allow future researchers to develop and test more sophisticated hypotheses about the complex interrelations between the multiple levels of influence and effective methods of cancer prevention and control.
Funding: This work was supported by: the predoctoral fellowship of SLP while at University of Texas School of Public Health (UTSPH Cancer Education and Career Development Program, National Institutes of Health, National Cancer Institute R25 CA05771) and the postdoctoral fellowship of SLP at WU (Barnes Jewish Hospital Foundation, Siteman Cancer Center Prevention and Control Program) and grant support of SWV’s time (National Institutes of Health, National Cancer Institute, RO1 CA97263 and RO1 CA112223). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.
The authors acknowledge Margaret Anderson for her assistance with identifying subject headings appropriate for the electronic database search and Stephanie Wuelling Russell for her help with data extraction.