Underuse of Radiation in Younger Women With Breast Cancer
We used a large, nationwide, employment-based MarketScan database that contains information on medical and outpatient prescription drug claims for employees and their spouses and dependents; the data represent claims from approximately 45 large employers and capture claims records from more than 100 payers. The University of Chicago's Institutional Review Board exempted this study for approval because all observations were deidentified.
We applied a previously published algorithm for claims data to identify incident cases of breast cancer. We first identified our study cohort as women aged 20 to 64 years who had at least one International Classification of Diseases, 9th Edition (ICD-9) diagnosis code indicating invasive breast cancer (ICD 9 CM 174.XX) and a procedure code of breast cancer surgery between January 1, 2004, and December 31, 2009. BCS was identified by common procedural terminology 4 and ICD-9 procedure codes. The earliest claims date indicating BCS was chosen as the index date. Patients who had mastectomy or both BCS and mastectomy claims concurrently were excluded. We then limited the study cohort to those who had continuous enrollment in health insurance 12 months before and after the index date to ensure that information obtained from claims in this 24-month window was complete. We excluded patients who had a breast cancer diagnosis more than 12 months before the index date, a prior history of breast cancer, RT before BCS, and mastectomy within 12 months of BCS. Lastly, we excluded patients with missing information on geographic region of residence or comorbidities and those with claims indicating distant metastasis. The final sample was 21 008 young women (Figure 1).
(Enlarge Image)
Figure 1.
Flow chart for data selection step. BCS = breast conserving surgery; ICD-9 = International Classification of Diseases, 9th Edition.
We identified RT using ICD-9 procedure, common procedural terminology, and revenue center codes based on published algorithms. We defined patients who were compliant as those who had any RT claims within 12 months of BCS. Other treatment-related variables included chemotherapy, use of staging imaging, axillary surgery, pre-BCS mammography screening, and inpatient vs outpatient BCS. Codes used to identify these variables are provided in Supplementary Table 1 (available online).
Other clinical variables included axillary lymph node involvement and whether the patient had a ductal-carcinoma-in-situ (DCIS) diagnosis within 12 months of the index date. Comorbidity was constructed using Klabunde's modified algorithm of Charlson comorbidity score.
We obtained patient age, geographic regions of residence, and whether the patient was the primary holder of insurance directly from the MarketScan database. We derived several patient characteristics variables using information provided in the data. First, we took advantage of the age information for the patients' dependents in the enrollment file to infer patients' family structure based on whether the family had a child of young age. Specifically, we classified families into four categories: families with at least one child aged less than 7 years, with no children aged less than 7 years but at least one child aged 7 to 12 years, with no children aged less than 13 years but at least one child aged 13 to 17 years, and those with no children or only children aged greater than 17 years. Our choice of cutpoints corresponds to the typical age ranges in nursery, elementary, middle/high school, and postsecondary education in the US education system. These categories reflect the child-care need, with the greatest need in families with at least one child aged less than 7 years. Second, we regrouped the insurance plan type variable into two categories based on whether patients were more likely to be tightly managed under the payment policy of their insurance: Health maintenance organization (HMO) or preferred provider organization (PPO) with capitation vs all other plan types. Third, we constructed a proxy variable to indicate whether patients may need to travel a long distance to receive RT by evaluating whether the geographic area (in terms of Census division) of the patient differed from that of her BCS provider.
We obtained variables of area characteristics by linking patients' county of residence to the 2004 to 2009 Area Resource File and measured the socioeconomic status (in quartiles) as the percentage of families with median household income below poverty level and percentage of adults with college degree. For health-care resources related to radiation therapy, we reorganized counties into health services area based on the modified health service area definition by the National Cancer Institute. Using health service area as the geographic unit that formed the patient's local market for RT, we included two supply-side variables in quartiles: density of hospitals with radiation therapy and density of radiation oncologists.
We used SAS version 9.2 (SAS Institute Inc. Cary, NC) for data management and STATA version 11.0 (StataCorp LP, College Station, TX) for statistical analyses. We used Pearson χ test to compare each covariable for women with RT within 12 months of BCS and those without and ran logistic regressions to examine factors associated with the receipt of RT. We applied three methods for model diagnosis, including 'linktest' in STATA, the Hosmer–Lemeshow goodness-of-fit test, and the Bayesian information criterion. The model specification that passed all three tests was chosen as the final model. In addition, we used perturbation analyses to assess potential impact of collinearity and misclassification. Next, we conducted stratified analyses by patients' age (aged 20–50, 51–55, 56–60, and 61–64 years) to determine whether certain covariables (eg, age of children) were only relevant to a specific age group. We also performed sensitivity analyses to evaluate the robustness of study findings across various subgroups. All statistical tests were two-sided.
Methods
Data Sources
We used a large, nationwide, employment-based MarketScan database that contains information on medical and outpatient prescription drug claims for employees and their spouses and dependents; the data represent claims from approximately 45 large employers and capture claims records from more than 100 payers. The University of Chicago's Institutional Review Board exempted this study for approval because all observations were deidentified.
Ascertainment of Study Cohort
We applied a previously published algorithm for claims data to identify incident cases of breast cancer. We first identified our study cohort as women aged 20 to 64 years who had at least one International Classification of Diseases, 9th Edition (ICD-9) diagnosis code indicating invasive breast cancer (ICD 9 CM 174.XX) and a procedure code of breast cancer surgery between January 1, 2004, and December 31, 2009. BCS was identified by common procedural terminology 4 and ICD-9 procedure codes. The earliest claims date indicating BCS was chosen as the index date. Patients who had mastectomy or both BCS and mastectomy claims concurrently were excluded. We then limited the study cohort to those who had continuous enrollment in health insurance 12 months before and after the index date to ensure that information obtained from claims in this 24-month window was complete. We excluded patients who had a breast cancer diagnosis more than 12 months before the index date, a prior history of breast cancer, RT before BCS, and mastectomy within 12 months of BCS. Lastly, we excluded patients with missing information on geographic region of residence or comorbidities and those with claims indicating distant metastasis. The final sample was 21 008 young women (Figure 1).
(Enlarge Image)
Figure 1.
Flow chart for data selection step. BCS = breast conserving surgery; ICD-9 = International Classification of Diseases, 9th Edition.
Identification of Radiation Therapy and Other Clinical Variables
We identified RT using ICD-9 procedure, common procedural terminology, and revenue center codes based on published algorithms. We defined patients who were compliant as those who had any RT claims within 12 months of BCS. Other treatment-related variables included chemotherapy, use of staging imaging, axillary surgery, pre-BCS mammography screening, and inpatient vs outpatient BCS. Codes used to identify these variables are provided in Supplementary Table 1 (available online).
Other clinical variables included axillary lymph node involvement and whether the patient had a ductal-carcinoma-in-situ (DCIS) diagnosis within 12 months of the index date. Comorbidity was constructed using Klabunde's modified algorithm of Charlson comorbidity score.
Patients and Area Characteristics
We obtained patient age, geographic regions of residence, and whether the patient was the primary holder of insurance directly from the MarketScan database. We derived several patient characteristics variables using information provided in the data. First, we took advantage of the age information for the patients' dependents in the enrollment file to infer patients' family structure based on whether the family had a child of young age. Specifically, we classified families into four categories: families with at least one child aged less than 7 years, with no children aged less than 7 years but at least one child aged 7 to 12 years, with no children aged less than 13 years but at least one child aged 13 to 17 years, and those with no children or only children aged greater than 17 years. Our choice of cutpoints corresponds to the typical age ranges in nursery, elementary, middle/high school, and postsecondary education in the US education system. These categories reflect the child-care need, with the greatest need in families with at least one child aged less than 7 years. Second, we regrouped the insurance plan type variable into two categories based on whether patients were more likely to be tightly managed under the payment policy of their insurance: Health maintenance organization (HMO) or preferred provider organization (PPO) with capitation vs all other plan types. Third, we constructed a proxy variable to indicate whether patients may need to travel a long distance to receive RT by evaluating whether the geographic area (in terms of Census division) of the patient differed from that of her BCS provider.
We obtained variables of area characteristics by linking patients' county of residence to the 2004 to 2009 Area Resource File and measured the socioeconomic status (in quartiles) as the percentage of families with median household income below poverty level and percentage of adults with college degree. For health-care resources related to radiation therapy, we reorganized counties into health services area based on the modified health service area definition by the National Cancer Institute. Using health service area as the geographic unit that formed the patient's local market for RT, we included two supply-side variables in quartiles: density of hospitals with radiation therapy and density of radiation oncologists.
Statistical Analysis
We used SAS version 9.2 (SAS Institute Inc. Cary, NC) for data management and STATA version 11.0 (StataCorp LP, College Station, TX) for statistical analyses. We used Pearson χ test to compare each covariable for women with RT within 12 months of BCS and those without and ran logistic regressions to examine factors associated with the receipt of RT. We applied three methods for model diagnosis, including 'linktest' in STATA, the Hosmer–Lemeshow goodness-of-fit test, and the Bayesian information criterion. The model specification that passed all three tests was chosen as the final model. In addition, we used perturbation analyses to assess potential impact of collinearity and misclassification. Next, we conducted stratified analyses by patients' age (aged 20–50, 51–55, 56–60, and 61–64 years) to determine whether certain covariables (eg, age of children) were only relevant to a specific age group. We also performed sensitivity analyses to evaluate the robustness of study findings across various subgroups. All statistical tests were two-sided.
Source...