These analyses utilized several data sources: (1) the 2000 and 2003 National Health Interview Surveys (NHIS) were used to examine utilization of CRC screening; (2) the Surveillance, Epidemiology and End Results (SEER)-Medicare File, for cancers diagnosed 1998–2002, was used for analyses of stage at diagnosis; (3) Medicare claims data, specifically the 2000 National Claims History (NCH or Carrier File), and Outpatient File, collected by Centers for Medicare and Medicaid Services (CMS) were used to create a measure of county CRC capacity; and (4) the Area Resource File (ARF) was used to create measures of county primary care physician (PCP) capacity, racial/ ethnic composition and socio-economic status (SES). This study was reviewed and approved by the Institutional Review Board of Partners HealthCare.
NHIS, conducted by the National Center for Health Statistics of the Centers for Disease Control and Prevention (CDC), was used for the analyses of CRC screening use. NHIS is a nationally representative household survey that collects information about demographic characteristics, chronic health conditions, health insurance, and health behaviors from the civilian, non-institutionalized U.S. population (www.cdc.gov/nchs/nhis.htm
). In 2000 and 2003, the NHIS included a Cancer Control Supplement, which asked respondents about the use of CRC screening.25
Because these analyses required merging NHIS with the area-level variables described below, all analyses were conducted through the Research Data Center of the National Center for Health Statistics.
SEER-Medicare data were used for analyses of stage at diagnosis. The SEER program collects information on all incident cancer cases for persons with cancer residing in SEER program areas (California, Connecticut, Hawaii, Iowa, Kentucky, Louisiana, New Jersey, New Mexico and Utah, rural Georgia, and the metropolitan areas of Detroit, Atlanta, and Seattle).26
SEER data include primary tumor site, stage at diagnosis, based on the modified American Joint Committee on Cancer (AJCC) staging system, and patient demographics and are linked to Medicare claims data by the National Cancer Institute. A restricted access version of these data was obtained so that the area-level characteristics could be appended.
Medicare claims, specifically the Carrier and Outpatient Files, for persons who had a screening sigmoidoscopy or colonoscopy were used to create aggregate measures of CRC screening capacity in the county where an individual resided. Data from the ARF was used to create a year 2000 county-level measure of PCP capacity and measures of area racial/ ethnic composition and SES. The ARF includes the number of physicians by specialty type and work setting in each county, the population by age in each county, the percentage of each county that is African-American and Hispanic and the median income for each county.
For the analyses of screening utilization, we used the NHIS data and included adults age 50 and older, who had not previously been diagnosed with colorectal cancer, whose race/ethnicity was reported as white, African-American or Hispanic, and who responded to questions regarding the use of CRC screening (n=23,229).
For the analyses of cancer stage, we used the SEER-Medicare data and included adults age 65 to 95, whose race/ethnicity was reported as white, black or Hispanic and who were diagnosed with modified AJCC Stage I, II, III or IV CRC as a single primary from 1998 to 2002 (n= 52,317).
In both the NHIS and SEER-Medicare samples, individual level variables included sex, age (10-year intervals), race/ ethnicity (white, African-American, Hispanic), marital status (married or living with a partner vs. other) and comorbidity. Comorbidity was measured slightly differently in the two datasets because of different data availability. In the NHIS dataset, participants were asked to endorse a variety of chronic conditions (e.g., arthritis, peptic ulcer disease, chronic lung disease). These were counted and categorized as 0, 1, 2, ≥ 3). For SEER-Medicare, claims data were used to calculate the modified Charlson comorbidity index (also categorized as 0, 1, 2, ≥ 3).27
Additional individual variables available in the NHIS included education (grade school or middle school, high school graduate or vocational school, some college, college graduate), prior history of cancer other than CRC, and health insurance (uninsured, Medicare with private supplemental insurance or private insurance, Medicare without supplemental coverage, and Medicaid or duel eligibility for Medicare and Medicaid). We also included health-care-seeking behavior, defined as having a usual source of care, evidenced by a visit to any health care professional in the past year, and the number of behavioral risk factors for CRC, including current cigarette use, heavy drinking (consuming 60 or more alcoholic drinks per month for men and 30 or more for women), and lack of regular exercise.28, 29
Additional individual-level independent variables in the SEER-Medicare dataset included whether an individual was of “low income” (based on eligibility for state assistance with Medicare premiums and co-payments), cancer type (colon, rectal), year of diagnosis, and an indicator of whether someone had non-continuous Medicare coverage or was in an HMO in the 13 months prior to diagnosis, in order to adjust for potentially incomplete data about chronic conditions for these individuals.
Area Variables CRC Screening Capacity
To create a measure of regional capacity for CRC screening, we first selected records from the 2000 Carrier and Outpatient Files that included a bill for screening sigmoidoscopy (Health Care Common Procedure Coding System (HCPCS): G0104, or Current Procedural Terminology (CPT) codes: 45305, 45308, 45309, 45315, 45320, 45331) or screening colonoscopy (HCPCS: G0105, G0121; CPT: 45380, 45384, 45385).30
There were a total of 1,198,597 unique Carrier and Outpatient records (multiple records actually make up a claim) based on patient identifier, procedure code and date of procedure for the 50 United States. Claims for the same procedure within 7 days were considered to be duplicates for the same procedures and the duplicate claim was deleted (n=8,976). Next, we identified the performing physician and zip code from the Carrier File, or the facility and county from the Outpatient File. Carrier Files with missing zip codes or with incorrect zip codes and Outpatient Files without county were deleted (n = 2,757), for a total of 1,186,864 records. Physicians who performed a colonoscopy or sigmoidoscopy were assigned to a county using zip code. If a zip code crossed county lines, it was assigned to the county with the greatest percentage of its population. Approximately 85% of zip codes were wholly contained within one county. Finally, we tabulated the number of sigmoidoscopies and colonoscopies performed in each county. We defined the screening capacity of the county as the number of colonoscopies and sigmoidoscopies performed in the county, per 100,000 residents age 50 and older. This measure was merged to the SEER and NHIS datasets by county.
Because PCPs are typically a gateway to CRC screening, we used data from the ARF for the number of PCPs in each county, defined as office-based physicians practicing the specialties of general practice, family practice, general internal medicine, or obstetrics/gynecology.31
We then calculated the number of PCPs per 100,000 residents age 50 and above for each county as our measure of PCP capacity. This measure was merged to the SEER and NHIS datasets by county. There were 844 counties included in the NHIS sample and 463 in the SEER-Medicare sample.
Racial/ ethnic Composition and SES
For NHIS, we used data from the ARF for the county racial/ethnic composition, defined as the percentage of the county that was African-American and Hispanic, and county SES, measured by median household income. The same measures were used, but at the census tract level, for the SEER-Medicare data.
Use of CRC Screening
Respondents to the NHIS were asked several questions about their use of CRC screening. Our measure for the use of CRC screening was whether an individual reported that they had ever been screened for CRC, using FOBT, sigmoidoscopy or colonoscopy. In a secondary analysis, we also examined whether an individual’s CRC screening had been done with either a sigmoidoscopy or colonoscopy.
Stage at Diagnosis
Stage at diagnosis was defined using the AJCC classification method that is available in the SEER-Medicare data (categorized as stage I, II, III or IV).
We used bivariate analyses to examine the associations of an individual’s race/ethnicity with demographic, area and outcome variables; chi-square tests and univariate linear regression models were used to determine statistical significance. To examine the effect of CRC screening capacity in an individual’s county of residence on individual screening, we constructed multi-level logistic regression models that accounted for clustering of individuals in the counties. To examine the effect of CRC screening capacity on stage at diagnosis, we constructed ordinal logistic regression models, clustered by county. In both sets of models, we included independent variables that we believed, a priori, could potentially affect the outcomes, based on our prior work and previous studies in the literature. The analyses of the NHIS also accounted for the survey sample weights. All area-level variables were modeled for a change of approximately 1 standard deviation (increments of 750 procedures per 100,000 individuals for CRC capacity, increments of 100 PCPs per 100,000 individuals, increments of 5% for the measures of racial/ ethnic composition, and increments of $20,000 for median household income). We tested correlations between the capacity measures and area racial/ethnic composition, and we tested for non-linear associations of both capacity measures with the outcomes by including a squared term for each measure in the models; in the model of CRC screening these quadratic trends were statistically discernable because of the large sample size but had a small magnitude of effect (e.g., odds ratio of 0.99) so, for ease of interpretation, the final models report only the linear trend. All analyses were done using SAS version 9.1 (SAS Institute, Cary, NC).
In order to examine whether area CRC screening or PCP capacity mediated any racial/ethnicity disparities in screening or stage at diagnosis, we built a sequence of models. First, we constructed models that included only individual-level variables. Specifically, the first CRC screening model included individual variables of age, sex, race, educational attainment, marital status, previous history of cancer, chronic conditions, behavioral risk factors for CRC, health insurance, usual source of care and whether the subject had a dental visit in the past year, and the year of the NHIS survey. The first stage at diagnosis model included the individual-level variables of age, sex, race, marital status, colon or rectal cancer, chronic conditions, whether an individual had ever been eligible for state buy-in insurance, and whether the individual was an HMO member or did not have Medicare for the 13 months prior to diagnosis, and year of diagnosis. For both outcomes, a second set of models added the measures of county CRC screening capacity, PCP capacity, and measures of the area racial/ ethnic and SES.