We used Medicare claims from fee-for-service beneficiaries in 10 eastern US states (NY, NJ, MD, DE, VA, NC, SC, GA, FL, AL) from a data linkage project that has been described previously.29 We accessed the data through the Centers for Medicare & Medicaid Services Chronic Conditions Data Warehouse (CCW) Virtual Research Data Center. The data included Medicare Part D files, carrier Standard Analytic Files, Chronic Conditions Data Warehouse (CCW) summary files, and the corresponding denominator files from 2010 to 2011. The Medicare Part D files contain brand and generic prescription drug names, strength, days’ supply, quantity dispensed, and service date. The carrier files contain noninstitutional provider claims for services covered under Medicare Part B, including International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes, place of service, Current Procedural Terminology codes, and dates of service. Medicare claims for laboratory services were linked to actual outpatient laboratory values processed by a national laboratory vendor in 2011 for Medicare beneficiaries. The CCW summary file provides flags for 27 chronic diseases based on diagnoses in Parts A and B Medicare claims from 1999 onward.30 Denominator files include patient demographic characteristics, birth and death dates, and information about program eligibility and enrollment.
We included older adults who met all the following criteria: enrolled in Medicare Parts A, B, and D between January 1, 2010, and December 30, 2011; had a diagnosis of diabetes in the CCW summary file in the 2 years before June 30, 2010; had at least one HbA1c lab result available between January 1, 2011, and June 30, 2011, and were 65 years or older at this HbA1c date; and were alive on December 31, 2011 (eFigure 1). If multiple HbA1c lab results existed for a patient, the first HbA1c lab was selected for analysis, and defined as the index date.
Diabetes medications were identified in the Medicare Part D files by searching, from 120 days before to 180 days after the index date, generic drug names for the following classes of medications: insulin, biguanides, sulfonylureas, thiazolidinediones, meglitinides, GLP-1 agonists, amylin analogs, DPP-4 inhibitors, and alpha-glycosidase inhibitors (see Appendix Table 5 for details). For patients on combination medications, we counted each component as a separate medication. Daily dose was calculated using medication strength, days’ supply, and quantity dispensed19 for all oral agents.
Patients were classified as potentially overtreated for diabetes if they had an index HbA1c value of less than 6.5% during the period from January through June 2011 and had filled any diabetes medications other than metformin in a pre-index medication regimen during the months leading up to this index HbA1c value. We chose this HbA1c threshold because it was substantially lower than the Choosing Wisely and AGS5 recommendations to avoid medications other than metformin when HbA1c values were less than 7.5%,31 and lower than the American Diabetes Association’s guidelines for glucose control for patients with diabetes.32 Patients were classified as potentially undertreated if their index HbA1c was greater than 9.0%, regardless of medication status, consistent with Healthcare Effectiveness Data and Information Set (HEDIS) measures since 1999.33 All others were classified as having acceptable treatment of their diabetes.
We defined medication deintensification among potentially overtreated older adults using a previously published method.19 Briefly, older adults on oral agents were defined as having had their regimens deintensified if they had medication fills with a lower dose than the index dose in the 1 to 180 days after the index date, or no refills of previously prescribed medications in the 22 to 180 days after the index date (see eFigure 1 in Sussman19 for an illustration). The index dose was defined as the medication dosage in the months leading up to the index HbA1c value used to define overtreatment, undertreatment, or neither. Medications filled in the 1–21 days after the date of the index dose were ignored for defining deintensification. We did this to account for blood draws or other visit-related work that occurs after the visit. Since HbA1c measurements are rarely repeated in under 3 months, this is unlikely to create any biases. Following prior work,19 older adults on insulin, pramlintide, or exenatide were defined as having deintensification if there were no medication refills in the 22 to 180 days after the index date. As the focus of this study was overtreatment and deintensification of therapy, we did not examine intensification of therapy for undertreated older adults, as many previous studies have examined this question.34
Covariates of interest obtained from the Medicare denominator files included baseline patient demographic characteristics, specifically age (65–70, 71–75, 76+), race/ethnicity (non-Hispanic white, Hispanic, non-Hispanic black, other),23 Medicaid dual eligibility, indication for Medicare eligibility (age vs. disability), gender, and urban location.35
, 36 We derived urban location from rural-urban commuting area codes based on ZIP Code of residence. The number of chronic conditions (0–1, 2–3, 4–5, 6+) was constructed from the CCW chronic disease flags at end of year for 18 conditions (see Appendix Table 6 for details). From carrier physician claims, we obtained the number of outpatient visits for evaluation and management (E&M) visits in 2010 based on Healthcare Common Procedure Coding System codes 992.xx, 993.xx, and 994.xx.
The associations between diabetes overtreatment and covariates of interest were summarized using means and standard deviations for continuous variables and frequencies with percentages for categorical variables. We compared rates of overtreatment to undertreatment and adequate treatment by groups, using χ2 tests for categorical variables and Kruskal-Wallis tests for continuous variables. To assess the association between patient characteristics and three-level treatment outcome, we used a multinomial logistic regression with the reference category as adequate treatment that adjusted for age, race/ethnicity, Medicaid dual eligibility, indication for Medicare eligibility (age vs. disability), gender, multimorbidity, number of outpatient visits, and urban location (Table 1).35
, 36 To account for multiple comparisons, we report 99% confidence intervals and used a two-tailed α = 0.01 to establish the statistical significance of tests for all analyses.
Among potentially overtreated individuals, we examined variables associated with treatment deintensification using means and standard deviations for continuous variables and frequencies with percentages for categorical variables. To test for differences between those who did and did not receive deintensification of treatment, we used χ2 tests for categorical variables and Kruskal-Wallis tests for continuous variables. We then examined whether deintensification differed based on race/ethnicity, dual eligibility, older age, sex, rurality (urban/rural), or multimorbidity, using multivariable logistic regression. Results are presented in both tabular and graphical form. The institutional review board of the Duke University Health System approved the study, and analyses were performed using SAS version 9.4 software (SAS Institute Inc., Cary, NC, USA).