Minnesota Managed Care Longitudinal Data Analysis. 2. Methods


As noted above, there are three components of the study: (1) the analysis of MSHO and MSC+ enrollment decisions for elderly dual eligibles; (2) the analysis of service use under MSHO and MSC+ by elderly dual eligibles; and (3) the comparison of elderly Medicare-only adults to the dual eligibles enrolled in MSHO and MSC+. This section describes the populations studied, data sources, measure construction, and statistical analyses performed.

2.1. Populations Studied

The sample for the principal analyses of utilization of dual eligibles is limited to 121,696 observations on full dual eligibles (having both Medicare Parts A and B and full Medicaid benefits) ages 65 and older who were consistently enrolled in either the MSHO or MSC+ program during any of the three years over 2010–2012 (excluding those who switched plans in a given year). Persons with intellectual or developmental disabilities or those who qualified for Medicaid as medically needy were excluded. Analyses were limited to adults in counties that offered at least one plan in each program in each year. A separate analysis was conducted of 25,162 dual eligibles who switched from MSC+ to MSHO.

One secondary set of analyses compares those dual eligibles to Medicare beneficiaries in Minnesota who were not dual eligibles as an additional comparison group. Sample size for this group ranged from 600,438 in 2010 to 612,052 in 2011 and 631,132 in 2012 (1,843,622 beneficiaries across three years).

2.2. Data

Data for the study were obtained from three sources for the State of Minnesota:

  • Dataset on dual eligibles containing fee for service claims, managed care encounters, and enrollment data.

  • Dataset on (Medicare-only) non-dual eligibles containing fee for service claims, and enrollment data.

  • Minimum Data Set (MDS) data containing nursing home assessments on dual eligible.

Claims and encounter data included information on inpatient, ED, outpatient, hospice, HCBS, and nursing home services. Acquisition of these data required two separate data use agreements with CMS and one with the State of Minnesota.

2.2.1. Dual Eligibles Data

These data were needed to address all research questions in the original Request for Proposal except for the single research question pertaining to non-dual eligible (Medicare-only beneficiaries). The data were in the possession of JEN Associates for supporting the State of Minnesota in its administration of the Medicaid program. JEN Associates made two datasets available--one at the person-month level and another at the claims/encounter-level.

2.2.2. Non-Dual Eligibles Data

These data were needed to address the research question pertaining to differences in individual characteristics and utilization between MSHO and the non-dual eligible population. These data were also provided by JEN Associates who produced both person-month and claims/encounter-level datasets.

2.2.3. Minimum Data Set Assessment Data

These data were used to assess differences in frailty across the MSHO and MSC+ populations at the time of nursing home entry. We had also sought to analyze State of Minnesota LTSS assessment data but learned in conversations with the state that some assessments were not routinely performed on all individuals and that some information of interest was not available on the assessments. Therefore, we decided not to analyze these data.

2.3. Measure Construction

We created person-year level files containing three years (2010-2012) of data from the person-month file provided by JEN Associates to create the following measures.

2.3.1. Enrollment

We created a dummy variable reflecting yearly enrollment, coded 1 if in the MSHO program throughout the year, and zero otherwise (that is, in the MSC+ program throughout the year).

2.3.2. Outcomes

We created nine measures of service utilization pertaining to any hospital inpatient care, outpatient ED use, long-term care nursing home use, overall physician use, PCP use, specialist use, HCBS (inclusive of assisted living facility use), assisted living, and hospice care. These measures were coded 1 if there was any use of each respective service, and zero otherwise, annually. We also created five count measures for levels of use reflecting the number of hospital inpatient stays, outpatient ED visits, overall physician visits, PCP visits, and specialist visits. Data on the level of long-term nursing home, HCBS, assisted living, and hospice use were complicated and construction of reliable count measures was beyond the scope of this project. Before using the count measures in regression analysis, we deleted extreme outliers at the far right tail of the distribution for the outpatient ED visit, overall physician visit, PCP visit, and specialist visit measures, which constituted only 0.05 percent of the overall sample for each measure. We did not delete any observations from the inpatient stay measure before regression analysis because no extreme outliers in the count of inpatient stays were observed in the data.

2.3.3. Individual Characteristics

Five dummy variables were created reflecting age groups (65-69 as the reference group, 70-74, 75-79, 80-84, 85-89, and 90+) and dummy variables for female gender and whether a person died during the year.

Five dummy variables were created for the following disability and medical conditions:

  • Mental illness (any diagnosis for Alzheimer's disease or dementia, chronic mental illness, depression, psychosis, or schizophrenia).

  • Neurological disability (any diagnosis for neurologic impairment or Parkinson's disease).

  • Physical disability (any diagnosis for physical impairment).

  • Sensory disability (any diagnosis for sensory impairment).

  • Other medical disability or chronic disease (diagnoses for selected medical disability or chronic diseases such as arthritis, chronic respiratory disease (chronic obstructive pulmonary disease [COPD], asthma, emphysema, or bronchitis), congestive heart failure, coronary heart disease, stroke, or diabetes).

2.3.4. Area-Level Characteristics

For regression analysis, we created and used either five area-specific measures or county fixed effects, to see if model estimates varied across these two formulations. The five area-specific covariates (all measured at the county level) were:

  • Number of PCPs per 1,000 population;
  • Percent of population 65+ who do not live in community;
  • Percent of population 65+ who live in community with others;
  • Percent of population 65+ with college education;
  • Percent of population 65+ who are married.


2.3.5. Minimum Data Set Measure Construction

In each year, we merged individuals in our analytic sample to nursing home resident assessments data from the MDS to identify newly admitted nursing home residents during the year. Specifically, a new nursing home admission was determined by the availability of an MDS assessment during a given year that is indicated as either an admission or a Medicare five-day or 14-day assessment. In addition, we looked retrospectively at the MDS data for each person to make sure the person had no prior nursing home use during the 100-day period before the date of admission to allow for "clearance" of prior nursing home use and establishing a new nursing home admission.

We focused on measures of cognitive impairment and limitations in performing five activities of daily living (ADLs): eating, toileting, transferring, bathing, and dressing. Changes for many MDS items, following the transition from MDS Version 2.0 to Version 3.0 in the last quarter of 2010, required that we separate results before and after 2010. Definitional differences between V2.0 and V3.0 posed a challenge for comparing MDS data over time between the two versions. For this descriptive analysis, the focus is on cross-sectional comparison of new nursing home admissions between individuals in the MSHO vs. MSC+ group, rather than differences over time.

2.4. Statistical Analyses

We analyzed the enrollment choice of MSHO and MSC+ enrollees and the impact of enrollment in a MSHO plan vs. in a MSC+ plan on the range of outcome measures for which data were available. We used logistic regression models for dichotomous outcome variables (enrollment choice and any use of each specific type of service). For count outcomes, we estimated negative binomial regression models that account for dispersion in the count data. In presenting the multivariate analysis results, we report odds ratios from logistic regression models and incidence rate ratios (IRRs) (which have a similar interpretation to odds ratios) from negative binomial regression models.

For both the enrollment choice and outcomes analyses we present descriptive statistics comparing MSHO and MSC+ enrollees to identify differences across the two groups. We performed multivariate regression analyses to determine the independent effect of the policy variable of interest (e.g., MSHO enrollment) on the outcome (e.g., any inpatient stay) after controlling for other individual and area-level characteristics. Regressions were performed on the 2010–2012 sample as a whole, controlling for calendar year effects, using 2010 as the omitted (reference) year in analyses. The regression models were run three times with the list of covariates differing each time. The three configurations of covariates were:

  • Beneficiary-level covariates only (Model 1 in tables).

  • Add specific area-level covariates to the Model 1 beneficiary-level covariates (Model 2).

  • Add county fixed effects (in place of specific area-level covariates) to the Model 1 beneficiary-level covariates (Model 3) (used for summary-level results when comparing the overall sample to urban and rural subgroups).

In multivariate analyses, we estimate a main model using the entire sample and the three versions listed above, but also estimate the same three models separately for urban counties and for rural counties. We anticipated that there would be some differences between urban and rural counties in terms of not only the populations, but also area-level factors. In reporting findings, we focus on Model 3, which provides the best goodness-of-fit measures across the three models. However, there is little difference in the findings across the three models, which yield a consistent assessment of the outcomes under MSHO relative to MSC+. The detailed model results for all outcomes are included in the Appendix.

For the MDS descriptive analysis, we assessed differences between MSHO and MSC+ nursing home admits in each year in physical and cognitive functions at the time of nursing home admission. We computed and compared the percentages of newly admitted nursing home residents with each of the select characteristics between residents in the MSHO group vs. those in the MSC+ group. We did this overall and stratified by age-sex groupings, where age is categorized into three broad groups (65-74, 75-84, and 85+) to ensure adequate sample size in each stratum. To increase sample size for robust descriptive statistics, we pooled data from 2011 and 2012 (based on MDS 3.0) for one set of analysis and data from 2008 and 2009 (based on MDS2.0) for a separate set of analysis. In addition, for 2008-2009 we added a third group--new nursing home admits among non-dual Medicare beneficiaries (hereafter also referred to as Medicare-only)--for comparison with new admits who were dual eligibles in the MSHO or MSC+ group.

There are likely to be unmeasured differences between MSHO and MSC+ enrollees that affect their health care utilization and, thus, have the potential to bias any comparison of outcomes under the two programs, making it difficult to assess the impacts of MSHO vs. MSC+. Therefore, we tested for potential effects of selection bias due to unobserved variables using a procedure developed for this purpose that presents upper and lower bounds for possible impact estimates had we been able to fully control for both observed and unobserved characteristics (Oster, 2015). In particular, we hypothesized that these omitted variables, which could include, for example, additional components of health and disability status, such as severity of chronic conditions and frailty, and family circumstances, such as marital status, living arrangements, and availability of informal caregivers, the potential to bias the estimates of the impacts of MSHO relative to MSC+ based on data available to this study.

View full report


"MNmclda.pdf" (pdf, 2.63Mb)

Note: Documents in PDF format require the Adobe Acrobat Reader®. If you experience problems with PDF documents, please download the latest version of the Reader®