Behavioral Disorders and Diabetes-Related Outcomes Among Massachusetts Medicare and Medicaid Beneficiaries

We conducted a retrospective cohort study using individual-level administrative data from the Medicare and Medicaid populations in Massachusetts. Medicare and Medicaid claims data were collected for Massachusetts residents in calendar years (CY) 2004 and 2005. The Medicare data used in this study came from the Centers for Medicare and Medicaid Services in the form of research identifiable files; the Medicaid data were obtained from the Massachusetts Medicaid Management Information System. The study was part of a larger analysis of treatment, service use, and expenditures for beneficiaries with behavioral disorders. Socioeconomic status data for beneficiaries' zip code-based community, median household income in 1999, and percentage of high school graduates were obtained from Census 2000. This study was approved by the University of Massachusetts Medical School Institutional Review Board.

We included beneficiaries who were at least 18 years old as of January 1, 2004. The study population was characterized according to their Medicare and Medicaid coverage. The Medicare-only group included beneficiaries who were enrolled in Parts A and B for at least ten months during both CY 2004 and 2005 and who were not enrolled in Medicaid. Beneficiaries with Medicaid only were enrolled in Medicaid for at least ten months during both years with no Medicare enrollment. Dually eligible beneficiaries were enrolled in Medicare Part A and Medicaid for at least ten months during both years. The purpose of the length-of-enrollment requirement was to ensure that the study population had stable and continuous health coverage. Beneficiaries with Medicare Advantage or who were residing in nursing homes for 90 days or more were excluded from the analysis.

All beneficiaries with type 2 diabetes were selected. They were identified by either one ICD-9-CM diagnosis on an inpatient claim or diagnoses on two outpatient claims (ICD-9-CM codes 250.xx, 357.2, 362.0, 362.01, 362.02, 366.41, and 648.0) in CY 2004. We could not link 2,263 individuals to a census area, and they were thus excluded from the study. As a result, the study included 106,174 individuals.

Behavioral disorders were identified by any one inpatient ICD-9-CM diagnosis or any two outpatient diagnoses in CY 2004. Mental disorders included schizophrenia and paranoid states (codes 295.x and 297.x), bipolar disorder (296.0, 296.1, and 296.4–296.7), depression or anxiety (codes 296.2, 296.3, 298.0, 300.01, 300.02, 300.4, 309.0, 309.1, 309.81, and 311.x), and other mental disorders (codes 298.1–298.4, 298.8–298.9, 300.1, 300.2, 300.3, 300.5–300.9, 301.x, 302.x, 306.x–308.x, 309.2–309.4, 309.82, 309.83, 309.89, 309.9, and 312.x–316.x). Mental disorders were categorized hierarchically, and individuals with more than one diagnosis were classified according to their most serious diagnosis. Schizophrenia or paranoid states were ranked as most serious, followed by bipolar disorder, then depressive or anxiety disorders, and finally other mental disorders. Individuals without any mental disorder diagnoses were the reference group. For example, an individual with diagnoses of schizophrenia and depression would be categorized in the schizophrenia and paranoid disorders group. Someone with bipolar disorder and depression would be categorized in the bipolar group.

Substance use disorders included alcohol abuse and dependence (codes 291.x, 303.x, 305.0, and 571.0–571.3) and drug abuse and dependence (codes 292.x, 304.x, 305.2–305.9, and 648.3). The reference group for any alcohol use disorder consisted of individuals without an alcohol use disorder, and the reference group was similarly constructed for any drug use disorder. A summary variable, any behavioral disorders, was created to indicate beneficiaries with any type of mental disorder or substance use disorder for the adjusted analyses (reference group: beneficiaries without any behavioral disorders).

Using ICD-9-CM and Current Procedural Terminology (CPT) codes, diabetes-related complications were identified from one inpatient or two outpatient claims during CY 2005 from the following categories: eye complications (ICD-9-CM codes 250.5 and 362.0), nephropathy (ICD-9-CM code 250.4), ischemic heart disease (ICD-9-CM codes 410–414), cerebrovascular disease (ICD-9-CM codes 431–436 and v1254), neuropathy (ICD-9-CM code 250.6), and lower-limb amputations (ICD-9-CM codes 250.8 and v497; CPT codes 27880, 27882, 27886, 28800, 28805, 28810, 28820, and 28825). The outcomes were analyzed separately and as a single measure (any complications). Diabetes-related hospitalization was assessed by any inpatient claims with any diabetes-related ICD-9-CM code—that is, diabetes and its complications—as the principal diagnosis.

Covariates included in the study came from two sources: the combined Medicare and Medicaid data and the Census 2000 data. The following variables were obtained from the Medicare and Medicaid data set: gender, race-ethnicity (whites, African Americans, Hispanics, and others), age, physical illness burden (represented by the Chronic Illness and Disability Payment System, CDPS [4]), preexisting diabetes-related complications (eye complications, nephropathy, neuropathy, lower-limb amputations, ischemic heart disease, and cerebrovascular disease) identified by ICD-9-CM codes, types of health coverage (Medicare only, Medicaid only, and dual eligibility), and frequency of physician visits. The CDPS was originally designed to allow the Medicaid program to predict health care expenditures for beneficiaries with ICD-9-CM-based diagnoses by assigning weights to diagnoses according to their expected use of resources (4). Therefore, a modified version of the CDPS score was used to approximate the severity of comorbid conditions by assigning more weights to more costly diseases, which are usually more severe. In estimating the CDPS scores, we omitted behavioral diagnoses from the assessment so that the measure represented physical illnesses only. Median household income and percentage of high school graduates in beneficiaries' zip code-based residential communities were obtained from Census 2000 data as proxy measures of the beneficiaries' socioeconomic status.

Unadjusted analyses were performed with chi square tests to compare the distributions of outcomes across behavioral disorder categories. We used logistic regressions to examine the associations between behavioral disorders on diabetes-related outcomes, adjusting for covariates. Two logistic models, one with the summary measure of any behavioral disorders and one with specific behavioral disorders, were performed to estimate the associations between adverse diabetes outcomes and any or specific behavioral disorders, respectively. Because treatment is likely to vary by practice or by provider and we were unable to identify specific providers across Medicaid and Medicare claims, we used Hospital Service Areas as a clustering variable to approximate small-area variations in care (5). We included robust variance estimates in the model to account for correlations in the patterns of care among patients within Hospital Service Areas (5). In these regression models, different population selection criteria were applied. Individuals with a particular previous complication were excluded from the analyses of the same complication as outcome. For example, people with previous eye complications (in 2004) were excluded from the analysis of eye complications as a 2005 outcome. No individuals were excluded from analyses of any diabetes complications and diabetes-related hospitalizations as outcomes. In adjusted analyses, age was categorized into four groups—younger than 55, 55–64, 65–74, and 75 and older—and CDPS scores were divided into quartiles. Levels of statistical significance were set at α=.05 for unadjusted analyses and adjusted analyses of any diabetes-related hospitalizations and at α=.01 for adjusted analyses of individual diabetes complication to adjust for the effects of multiple comparisons. All statistical analyses were performed with Stata, version 10.