Geographic variation and effect of area-level poverty rate on colorectal cancer screening

Data from the 2006 Missouri BRFSS were used to assess the geographic variation and the effect of area-level poverty rate on CRC screening. The BRFSS is a survey of health risk factors and is sponsored by the Centers for Disease Control and Prevention (CDC). It is a standardized, random-digit-dialed telephone survey of the noninstitutionalized adult U.S. population carried out in all 50 states and the District of Columbia. In 2006, 5,391 Missourians age 18 or older were interviewed. The response rate of the 2006 Missouri BRFSS was 58.5%. To account for the complex sampling design, BRFSS data were weighted to adjust for the unequal probability of selection, differential nonresponse, and possible deficiencies in the sampling frame. We used a three-level design with individuals nested within self-reported five-digit ZIP code areas (ZIP5), which, in turn, were nested within three-digit ZIP code areas (ZIP3) that ignore the last two digits of ZIP codes.

In multilevel studies, a frequent concern is which geographic level should be considered [39] in order to target particular geographic areas. Inappropriate use of geographic levels can lead to the "Modifiable Areal Unit Problem" (MAUP) [40]. MAUP results from the use of artificial geographic units which were constructed from a continuous geographic area and may lead to the misinterpretation of results. A three-level multilevel design reduces the likelihood of MAUP bias to some degree [40].

The Institutional Review Board of the Missouri Department of Health and Senior Services (MDHSS) and Washington University reviewed the study protocol and determined it to be exempt. The original database provided by MDHSS did not contain identifiable information of study subjects.

According to the American Cancer Society's (ACS) guidelines [2], persons age 50 or older who are at average risk for CRC should have an annual fecal occult blood test (FOBT), a sigmoidoscopy every five years, or a colonoscopy every ten years. Because the BRFSS questions did not allow for the separation of colonoscopy and sigmoidoscopy use, we defined CRC screening adherence as having a FOBT in the past year and/or having a sigmoidoscopy/colonoscopy in the past five years for persons age 50 or older. We recognize that persons who reported having a colonoscopy 6–10 years prior to their interview would be misclassified as not adhering to ACS guidelines, thereby underestimating overall CRC screening rates. Because of the differences between FOBT and use of endoscopy (colonoscopy and sigmoidoscopy), we also performed the analysis separately for each test.

We used the percentage of the population living below the US federal poverty line from the 2000 census as a measure of area socioeconomic position. The poverty rate appears to be a measure that is robust across various diseases and levels of geography; it has policy implications, and is comparable over time [39].

We linked self-reported ZIP codes from the BRFSS to the 5-digit and 3-digit Zip Code Tabulate Areas (ZCTA5 and ZCTA3) from the 2000 census to obtain ZIP code-level poverty rates [41]. The ZCTA5 and ZCTA3, constructed from census blocks, are close approximations to 5-digit and 3-digit ZIP codes [42]. The ZCTA5- and ZCTA3-level poverty rates were categorized into less than 10%, 10–19% and 20% or higher to allow for nonlinear effects.

Data about the individual-level factors were obtained from the 2006 Missouri BRFSS and selected based on their associations with CRC screening [11, 14, 18‐22]. We considered 18 individual-level covariates in six groups as potential factors by which area-level poverty would be associated with CRC screening (Table 1): (1) demographic characteristics (age, gender, race/ethnicity, marital status, education, employment status and family income); (2) personal health condition (body mass index [BMI], history of one or more health conditions [asthma, diabetes, heart attack, coronary heart disease, and stroke], and history of depression and/or anxiety); (3) access to medical care (health insurance coverage, having a primary care physician, and having a routine check-up); (4) social support; (5) health-related behaviors (current smoking status, alcohol use, and participation in physical activity); and (6) self-rated health status.

Age was dichotomized as age 50–64 and 65 or older. Race/ethnicity was categorized as non-Hispanic white, non-Hispanic African American, and Hispanic/others. Marital status was grouped as married or not. Education level was categorized as college or higher level or less than college level. Employment status was categorized as employed, not employed, retired or refused. Household income was categorized in three groups: less than $25,000, $25,000–$50,000, and $50,000 or more. BMI was analyzed in three categories: <25.0, 25.0–29.9, and 30 or higher. Having ever been told by a doctor, nurse or other health professionals to have one or more of five chronic diseases, including diabetes, heart attack, angina or coronary heart disease, stroke and asthma, was considered as having a chronic disease history. Anxiety/depression history was defined as ever being told by a doctor or other health care providers to have an anxiety or depressive disorder. Having health insurance and having a primary care physician were both categorized as "yes" or "no." Routine medical check-up was categorized as "within one year," "two to five years," "more than five years" and "don't know." Always or usually needing emotional social support was compared to sometimes or never. We categorized smoking status as never smoked, former smoker, and current smoker based on responses to the following two questions: "Have you smoked at least 100 cigarettes in your entire life?" and "Do you now smoke cigarettes every day, some days, or not at all?" A never smoker was an individual who answered no to the first question. A current smoker was defined as an individual who responded yes to the first question and "every day" or "some days" to the second question. A former smoker is an individual who answered yes to the first question and "not at all" to the second question. Alcohol use was categorized as being a current heavy drinker or not. Heavy drinking was defined as having more than two drinks per day for men and having more than one drink per day for women. Persons who reported to have participated in physical activity or exercise during the past 30 days other than their regular job were considered as having been physically active. Self-rated health status was grouped into two categories: "good or better health" and "fair or poor health."

Data were analyzed using three-level logistic regression models. A total of 3,022 participants age 50 or older (Level one), were nested within 1036 ZIP5 areas (Level two), which, in turn, were nested within 25 ZIP3 areas (Level three). The complex sampling design in BRFSS was taken into account in the analysis. Models were weighted by a normed weight on the basis of the final weight variable that was calculated by the CDC. This normed weight was the multiplication of the BRFSS-final-weight by the ratio of participants age 50 or older in Missouri [43]. All models were fitted by second order penalized quasi-likelihood (PQL) estimation using the MLwiN 2.02 software [44]. We found no evidence of extra binomial variation using a chi-square test in an empty model without any covariates or poverty rate, suggesting that the logistic model was appropriate. Fixed effects of poverty rate and random effects of CRC screening use across ZIP5 and ZIP3 were estimated for all models. To quantify the random effect of CRC screening, random intercept models were constructed in all analyses. The deviance information criterion (DIC) was used to evaluate the model fit. Smaller DIC values mean better model fit. Intraclass correlations (ICC) were computed to estimate the proportion of geographic variation across the ZIP5 and ZIP3 areas of the total variation, including variation across and within the ZIP5 and ZIP3 areas. The variation at level one was fixed at π ²/3 (approximately 3.29) [45]. Because of the limitation of the ICC for binary response variables when measuring geographic variations and comparing fixed effects of area-level variables, we applied the methodology developed by Larsen and colleagues to compute the median odds ratio (MOR) and the 80% interval odds ratio (IOR) [46]. The MOR was calculated to quantify the variation of CRC screening across ZIP5s and ZIP3s using the equation: M O R = exp ( Z 0.75 × 2 × V a r ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemyta0Kaem4ta8KaemOuaiLaeyypa0JagiyzauMaeiiEaGNaeiiCaa3aaeWaaeaacqWGAbGwdaWgaaWcbaGaeGimaaJaeiOla4IaeG4naCJaeGynaudabeaakiabgEna0oaakaaabaGaeGOmaiJaey41aqRaemOvayLaemyyaeMaemOCaihaleqaaaGccaGLOaGaayzkaaaaaa@447A@. Where Var is the variance at ZIP5- or ZIP3-level, and Z_0.75 is the value of standard normal distribution at 75% percentile (equals 0.6745). The "exp" is the exponential function. The MOR is always equal to or greater than one, and a larger MOR value indicates a larger geographic variation.

The IOR was calculated to quantify the fixed effect of area poverty rate using equations: I O R U P P E R = exp ( β + Z 0.90 × 2 × V a r ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaKKaem4ta8KaemOuai1aaSbaaSqaaiabdwfavjabdcfaqjabdcfaqjabdweafjabdkfasbqabaGccqGH9aqpcyGGLbqzcqGG4baEcqGGWbaCdaqadaqaaiabek7aIjabgUcaRiabdQfaAnaaBaaaleaacqaIWaamcqGGUaGlcqaI5aqocqaIWaamaeqaaOGaey41aq7aaOaaaeaacqaIYaGmcqGHxdaTcqWGwbGvcqWGHbqycqWGYbGCaSqabaaakiaawIcacaGLPaaaaaa@4CEA@ and I O R L o w e r = exp ( β + Z 0.10 × 2 × V a r ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaKKaem4ta8KaemOuai1aaSbaaSqaaiabdYeamjabd+gaVjabdEha3jabdwgaLjabdkhaYbqabaGccqGH9aqpcyGGLbqzcqGG4baEcqGGWbaCdaqadaqaaiabek7aIjabgUcaRiabdQfaAnaaBaaaleaacqaIWaamcqGGUaGlcqaIXaqmcqaIWaamaeqaaOGaey41aq7aaOaaaeaacqaIYaGmcqGHxdaTcqWGwbGvcqWGHbqycqWGYbGCaSqabaaakiaawIcacaGLPaaaaaa@4DD4@. Where β is the regression coefficients of area-level poverty rate, Z_0.90 and Z_0.10 are the values of standard normal distribution at 90% and 10% percentiles, respectively (equals ± 1.2816) and Var is the variance at the corresponding ZIP-level poverty rate. The range of IOR values also reflects the degree of variation between areas. A substantial contribution of the poverty rate to the geographic variation of CRC screening is present when the IOR does not contain the value of one.

To examine the associations between individual-level factors and CRC screening, a multilevel logistic regression analysis, involving all individual-level factors and area-level poverty, was used to estimate adjusted odds ratios and their 95% confidence intervals. Then, we constructed nine models to examine the role of area-level poverty rate on CRC screening. First, an empty model without any fixed effects was constructed to calculate the MORs at level 2 and 3. Second, we determined the unadjusted association of ZIP5- and ZIP3-level poverty rate with CRC screening use. Third, we assessed the potential mediating effect of each of the six hypothesized pathways on the association between ZIP5- and ZIP3-level poverty rates and CRC screening by constructing a separate model which included a set of individual variable and ZIP5- and ZIP3-level poverty rates. Finally we also fitted a model with ZIP5- and ZIP3-level poverty rates and all individual-level variables. Reduction in odds ratios and MOR for the ZIP5- and ZIP3-level poverty rate relative to unadjusted analysis was used as evidence for mediation. We also reran the analyses with only individual-level covariates significantly associated with CRC to reduce the likelihood of collinearity among these covariates.

The multivariable multilevel models adjusted for all individual covariates and area-level poverty rate showed that people who were retired, with higher BMI and anxiety/depression history had higher odds of CRC screening (Table 2). In contrast, non-African American minorities, people with a lower education level and lack of access to care had lower odds of CRC screening. People with mid-low household income ($25,000 to 50,000) also had slightly lower odds of having CRC screening. The strongest individual-level indicator for being non-adherent to CRC screening is lack of access to medical care specifically lack of health insurance coverage (OR = 0.57, 95% CI: 0.35–0.90), having no primary care physician (OR = 0.48, 95% CI: 0.32–0.74) and not having a routine check-up (2–5 yrs: OR = 0.41, 95% CI: 0.32–0.54; >5 yrs: OR = 0.12, 95% CI: 0.05–0.27).

Table 2 also shows differences in the associations of the individual covariates with FOBT and endoscopy use. While older Missourians, African Americans, those overweight or obese, those with a history of anxiety/depression, and those considered to be heavy drinkers were more likely to report having had an endoscopy in the past five years; non-African American minorities, persons with a high school or lower education, and those without health insurance coverage or a primary care physician were less likely to report having had an endoscopy within the past five years. However, none of these covariates were associated with FOBT use in the past year. Only Missourians without a routine check-up were significantly less likely than those with a routine checkup to report having FOBT use in the past year.

At the contextual level, increased second- and third-level area poverty rates were associated with lower CRC screening (Table 3). All nine models indicated significant geographic variation across ZIP5 areas, but not across ZIP3 areas. All nine ZIP5-level MOR values were similar (about 2.0), suggesting that the geographic variation was relatively large and indicating that, for subjects with the same individual covariates, persons living in areas with high poverty rates had about two times of median odds of not following CRC screening guidelines compared with persons living in areas with lower poverty rates. The MOR is related to the fixed-effects odds ratios (OR) for area-level variables. While the description of the adjusted OR of an area-level variable is fixed across the geographic areas, the MOR describes the variation across the geographic areas accounting for all area- and individual-level characteristics in the multivariable model. More details about the MOR are described by Larsen and Merlo [46].

The results for CRC screening closely followed the results for endoscopy use (Table 4). However, results were different for FOBT use (Table 5). There was no association between area-level poverty rate and FOBT use within the past year. The geographic variation in FOBT use was substantially larger across ZIP5 areas than for endoscopy use.