This cohort of nursing homes was defined utilizing the Online Survey Certification and Reporting (OSCAR) database and the Residential History File (RHF). The OSCAR data are available from the Centers for Medicare and Medicaid Services (CMS) and are based on data collected from each Medicare/Medicaid certified nursing home during annual survey inspections. The RHF is a unique data resource built using Medicare enrollment data, Medicare claims data, and Residential Assessment Instrument Minimum Data Set (MDS) data.12
The RHF can be used to track individuals as they move through the long-term care system, including between different care settings and different care types (e.g., hospice). These data were obtained through a data use agreement (DUA) with CMS.
Over the 7-year study period, 131,458 nursing home observation years were identified. Nursing home observation years were excluded if the nursing home was hospital-based (n=15,464), had fewer than 30 beds or more than 500 beds (n=3337), had fewer than 3 years of observation (n=11,607), had fewer than five deaths in any year (n=17,167), or was missing data on covariates (n=11,776). This resulted in 72,107 nursing home observation years among 10,759 nursing homes.
The dependent variable was defined as the number of CNA minutes per resident per day (MPRD), a measure of nursing home direct-care staffing. Facilities report the number of CNA hours during the 2 weeks prior to their annual survey. CMS converts the number of hours into full-time equivalents (FTEs; based on a 35-hour work week) and this is reported in the annual OSCAR data. Nursing home staffing data derived from OSCAR have been used in numerous previous studies.13
For this study, we converted the FTEs back into hours, by multiplying by 35, and divided the total number of CNA hours by the number of residents in the facility (also drawn from the OSCAR) to arrive at the CNA hours per resident day (HPRD). For ease of interpretation, we further transformed HPRD into MPRD by multiplying by 60. We also cleaned this variable when the FTEs reported were implausible (when total FTEs were 995 or higher or if there were more than three times the number of CNAs reported than the number of beds in a facility). If data were missing in a particular year, we imputed the values by randomly choosing a value from the quartile of nursing homes having the most similar values (to the “missing value” nursing home) in the subsequent year. Less than one-half of 1% of facilities had missing values each year (and in most years it was between 0.1% and 0.2%).
We operationalized the introduction of hospice services as a nursing home transitioning from having less than 1% of its total resident days in a calendar year covered by hospice to more than 1%. The RHF was used to identify the number of nursing home days for all residents in each facility in the calendar year as well as the number of those days that were hospice days. The proportion of days that were hospice was calculated using these two counts. A dichotomous variable was then created to indicate when a nursing home reached the 1% threshold that signified “adding hospice.” Hospice volume was operationalized by totaling the number of resident hospice days in the nursing home in a calendar year. For ease of interpretation, this variable was standardized to 1000 nursing home hospice days.
We controlled for numerous nursing home attributes including the total number of resident days per calendar year, and for other time varying nursing home characteristics including a yearly indicator of whether more than two thirds of a nursing home's hospice residents were cared for by for-profit hospices, the proportion of residents on Medicaid, the proportion of residents on Medicare, occupancy rate, resident case mix, nurse staffing, existence of a special care unit, and existence of a physician extender (nurse practitioner or physician's assistant). To control for secular trends, we also included indicator variables for each year. We included two measures of resident case mix, one based on all MDS admission assessments (i.e., the case mix for all newly admitted residents in a year) and one based on all MDS annual assessments (i.e., the case mix for long-stay residents). We used the resource utilization group (RUG III) system, which classifies residents into categories based on their estimated resource utilization. This approximates the relative staff time associated with caring for the average resident in each group. Thus, the higher the case mix score, the more severe the average acuity profile of the residents in a nursing home. These measures were standardized, so that the regression results reflect the influence of a one standard deviation change in case mix.
To examine the influence of hospice competition, we included a variable indicating the number of hospices in the nursing home's county providing services in nursing homes each year. We identified hospice providers who provided services to nursing home residents each year as identified on Medicare claims in the RHF. We then grouped residents by their nursing home's county and added up the unique number of hospices that were providing care to those residents.
We performed panel multivariate regression analysis using a facility fixed-effects model. A panel fixed-effects model controls for unobserved time-invariant nursing home characteristics that may be correlated with CNA MPRD, and allows for a difference-in-difference model causal interpretation of our results. We used the XTREG (cross-sectional time-series linear regression) procedure available in Stata statistical software version 10, which fits regression models to panel data.14