Our retrospective analysis explored changes in CRC testing during a 5-year period that included the transition of Medicaid enrollees into CCOs (2012–2013) and the implementation of Medicaid expansion (2014). Our analysis was conducted and refined April 2016 through June 2018. The study was approved by the Oregon Health & Science University Institutional Review Board with a waiver of informed consent (IRB #8865).
Setting
Oregon is noted for its experimentation with Medicaid enrollment policies [28, 29]. In 2012, Oregon initiated assignment of Oregon Health Plan Members (the state’s Medicaid Program) into 16 regionally-based CCOs. CCO characteristics are described in detail elsewhere and summarized here [30,31,32,33]. CCOs are similar to ACOs in that they are responsible for providing coordinated health care services to patients in their region while controlling costs [34]. CCOs provide coverage for than 90% of Oregon’s Medicaid population; those not enrolled in a CCO receive care through the state’s Medicaid fee-for-service program for reasons related to special health needs (e.g., medically fragile children). Each CCO is governed locally by a board consisting of healthcare providers, community members, and other stakeholders. CCOs operate within a set budget based on the number of enrollees, with fixed annual percentage increases in funding [30]. As detailed in Additional file 1, Oregon’s CCOs vary by geographic region, size (range: 11,347 to 228,263 enrollees) and racial/ethnic composition (range: 6.2 to 33.3% Hispanic). CCOs include both non-profit (n = 9) and for-profit (n = 7) entities.
CCOs are accountable to the state through the tracking of multiple quality incentive measures that encompass domains ranging from preventive care to outpatient and emergency department utilization [35]. CCOs that meet improvement targets or benchmarks are eligible for annual performance bonuses from the state [30]. CRC screening has been an incentive measure since the first year of the CCO program, with reporting initiated in 2013 [35, 36].
Data collection and analysis
Data source
Administrative claims and enrollment data for Medicaid enrollees were obtained from Oregon’s Health Systems Division for a five-year period from January 1, 2010 through December 31, 2014. Claims data include all healthcare encounters that generated a billing claim for enrolled members over the study period. Claims data have been widely used to understand cancer screening patterns in diverse insured populations [9, 37].
Identification of eligible patients
We applied the same inclusion and exclusion criteria for each calendar year to generate the analytic sample of eligible Medicaid members: aged 50 to 64 years, not dually insured by Medicare, had a valid zip code, alive for the entire study period, and continuously enrolled in a CCO (defined as enrolled for at least 11 of 12 months). We excluded individuals enrolled in Medicare since we did not have access to Medicare claims data. Consistent with prior analyses, we also excluded enrollees that had resided in more than two counties during the entire study period [13] and those with end-stage renal disease, a terminal illness that would preclude clinicians from recommending cancer screening [9, 38]. We excluded beneficiaries with a history of CRC or total colectomy to better ensure testing eligibility [9]. A total of 132,424 unique Medicaid enrollees met inclusion criteria over the study period (see Additional file 2).
Primary outcome measures: CRC testing and modality used
We assessed CRC testing by colonoscopy, flexible sigmoidoscopy, or FOBT/FIT to be consistent with USPSTF guidelines active during the study window [3, 39]. In the rare event of multiple CRC testing modalities billed on the same day of service, we recorded modality based on the most invasive test received (i.e., colonoscopy > flexible sigmoidoscopy > FOBT/FIT). CRC testing was identified by International Classification of Diseases, 9th Edition, Clinical Modification (ICD-9-CM), Current Procedural Terminology (CPT), or Healthcare Common Procedure Coding System (HCPCS) codes, which are summarized in Additional file 3. For FIT/FOBT, we examined non-specific codes (i.e., 82,271, 82,272, 82,273) to explore how often they occurred concurrently with a CRC screening test-specific code and discovered that usage of non-specific codes decreased in a stepped fashion from 27.5% in 2010 to 11.0% in 2014. Because this decrement may be related to coding improvement and not differences in testing behaviors, we kept non-specific procedures in the analysis. We included both screening and diagnostic billing codes for colonoscopy in our analysis, consistent with prior studies [9, 40].
Statistical analysis
First, we used descriptive analyses to determine the number of eligible Medicaid enrollees, their demographic characteristics, and the observed rates of CRC testing annually.
We used multivariate linear regression at the patient-level to determine the probability of CRC testing for each calendar year with reference to the baseline year (2010). Our models controlled for patient-level characteristics including age, gender, race/ethnicity (White, African-American, Hispanic, Asian/Pacific Islander, American Indian/Alaska Native, other/unknown), urbanicity (urban, rural/frontier), chronic disease risk, and use of primary care within the calendar year. We categorized urbanicity using the ZIP-code version of the Rural-Urban Commuting Areas (RUCA) taxonomy based on population density, urbanization, and daily commuting patterns: urban (50,000 or more) versus rural (2500-49,999) and frontier (< 2500) [41]. We computed Chronic Illness Disease Payment System (CDPS) indicators from claims data to adjust for chronic disease risk [42]. Use of primary care within the calendar year was determined from claims based on date of service and a standard set of CPT codes [43]. CCO fixed effects controlled for possible clustering of CRC testing outcomes at the CCO level, and calendar quarter indicators controlled for seasonality.
In addition to regressing on overall CRC testing, we also fit similar modality-specific models - i.e., where the outcome is colonoscopy, flexible sigmoidoscopy, or fecal testing. We also evaluated our selection of a linear probability model over logistic regression, a common alternative approach for binary outcome variables. Predicted probabilities of CRC testing using the linear model were found to be highly correlated with predicted probabilities from a logistic regression (Pearson’s correlation coefficient 0.961, p < 0.001). When concordance between approaches is high, the linear model has the advantage of more intuitive interpretation of the coefficients of the independent variables as marginal effects.
Finally, we descriptively examined the annual percentage of enrollees in each CCO who received CRC testing overall and by modality. All statistical analyses were conducted in R version 3.2.2. To protect confidentially, we de-identified CCOs when reporting data that was not publically available.