Patients with newly diagnosed cancer or patients undergoing all or part of their initial treatment at JH SKCCC were identified through the Johns Hopkins Hospital (JHH) cancer registry. All accruals to cancer clinical trials, including therapeutic and non-therapeutic studies, were recorded by the JH SKCCC Clinical Research Office. This study was designated as exempt by the institutional review board of the Johns Hopkins School of Medicine.
This study included analytic cancer cases and accruals to clinical trials enrolled from 2005 to 2007. Year of diagnosis for each case was defined based on the composite medical record, with preference going to a pathology report of cancer for date of diagnosis. Cancer center analytic cases are defined by the American College of Surgeons as initially diagnosed or receiving all or part of the first course of treatment within the center.15
Study subjects were residents of the continental United States at diagnosis. This analysis comprised 17,637 analytic patients and 5068 accruals.
Level of clinical trial participation was measured by the ACR and defined as the number of accruals divided by the number of cancer cases diagnosed in the same period and population subgroup. JH SKCCC cancer cases and clinical trial accruals were aggregated by case characteristics of age, race, sex, place of residence, and cancer site. Because the portfolio of therapeutic and nontherapeutic clinical trials differed substantively, results were presented by clinical trial type separately.
Numbers of clinical trial accruals among Hispanic individuals were small and thus not considered in this Poisson, multivariate analysis, which requires sufficient and non-zero sample size in most cells. During the 5-year period, the JH SKCCC registered 163 Hispanic patients and 79 accruals; nevertheless, these patients are included in the race (white, black, and other races) analysis. Place of residence and race were the independent variables of interest.
Place of residence was categorized as Baltimore City; non–Baltimore City catchment area, consisting of 57 surrounding counties; or non–catchment area (i.e., all other continental United States counties of residence). The JH SKCCC catchment area is a SaTScan-determined geographic cluster of any cancer case seen within the center (1998–2002).11,12
County of residence was maintained as state and county/city name in the JHH cancer registry, which were converted to Census Bureau Federal Information Processing Standards (FIPS) codes. Zip codes of residence were converted to county FIPS codes using Market Planner Plus software (Solucient) for patients whose addresses were incomplete. Cancer cases and clinical trial accruals with county address of residence unresolved were coded as “unknown” (cancer cases, N = 741; clinical trial accruals, N = 132).
Race was abstracted by the JHH cancer registry staff from the medical record. Most JH SKCCC patients could be categorized as white or black (97%); all persons of non-white or non-black race were coded as “other races.”
Cancer case and clinical trial accrual characteristics of age (< 20, 20–64, > 64 years), sex (male, female), and cancer site (brain, breast, gastrointestinal, hematologic cancers, prostate, upper aerodigestive, and all other sites) were abstracted from the medical record. In therapeutic accrual data, 3 patients are missing race information, 94 are missing age, and 1 is missing the cancer site. Among nontherapeutic accruals, 8 are missing race, 61 are missing age, and 1 is missing cancer site. Observations with missing information were excluded in multivariate analysis.
Cancer site groupings were based on the JH SKCCC organizational structure and reflect programmatic structures within JH SKCCC. During Poisson regression analyses, cancer sites were further aggregated into 3 classes based on the level (high, medium, and low) of the unadjusted therapeutic trial accrual ratio. The first cancer site grouping consisted of hematologic malignancies; the second, prostate and gastrointestinal; and the third, all others. This strategy also had the advantage of allocating approximately one third of the therapeutic trials to each group.
Poverty level was assigned based on county of residence using FIPS code and defined as percentage of individuals living at or below 100% poverty in 2003.16
The authors presented poverty level by county quartiles (≥16.2; 12.6–16.1; 9.8–12.6; ≤ 9.8%) and unknown in bivariate analyses. In Poisson analyses, binary categories of poverty of 16.2 or more, and poverty less than 16.2, including poverty unknown, were used to minimize the number of combinations during multivariate analysis.
This study measures the accruals among JH SKCCC patients. Poisson regression modeling of clinical trial accruals adjusted to JH SKCCC cancer cases (i.e., offset) was used to estimate the ACR for therapeutic and nontherapeutic clinical trials. Results of Poisson regression modeling did not differ from zero-inflated models was presented here.
A clinical trial ACR “relative risk” (RR) was calculated. The ACR RR compared the ACR for various subpopulations to the reference group. Sub-populations were based on age, race, sex, cancer site, place of residence, or county poverty. The ACR RR is the ratio of 2 ratios: the subpopulation ACR and the reference group ACR. The authors hypothesized an interaction of residence in Baltimore City and participation in clinical trials by African Americans; that is, this grouping of clinical trial accruals, adjusted for patients from the same grouping, differs from any other race–geography grouping. This was tested explicitly and was not significant, so the final multivariable models considered all independent variables and covariates without interaction terms. Statistical significance of subpopulation differences was based on analysis of variance (ANOVA).
Because poverty and minority race are closely associated,17
the authors calculated the variance inflation factor (VIF) statistic, which measures the magnitude of multicollinearity, for the final regression model. The VIF was less than 2.8 for all predictors, well within the judgement of little evidence for multicollinearity.18
Statistical software R was used for all analyses.