Methods
Design and Data Collection
This was a cross-sectional ecologic study aimed at exploring relationships between aggregate residential characteristics and QMs. Data were retrieved from the 2019 annual archives found in the CMS provider data catalog on nursing homes, including rehabilitation services.11 The dataset provides general residence information, such as ownership, number of beds, number of residents, and location, as well as residence quality metrics, such as QMs, staffing data, and inspection data. Residence characteristics and 4-quarter averages of QMs were retrieved and used as cross-sectional data. The data used are from 15,420 residences across the United States. Nursing homes located in Guam, the US Pacific Territories, Puerto Rico, and the US Virgin Islands, while supported by CMS and included in the dataset, were excluded from the study due to a severe absence of QM data.
Dependent Variables
We investigated 5 QMs that were averaged across the 4 quarters of 2019. The QMs used as dependent variables were percentage of residents who displayed depressive symptoms (depression), percentage of residents who were physically restrained (restraint), percentage of residents who experienced 1 or more falls resulting in a major injury (falls), percentage of residents who received antipsychotic medication (antipsychotic medication), and percentage of residents who received anti-anxiety or hypnotic medication (anti-anxiety medication).
A total of 2471 QM values were unreported across the 5 QM analyzed: 501 residences did not report depression data; 479 did not report restraint data; 477 did not report falls data; 508 did not report antipsychotic medication data; and 506 did not report anti-anxiety medication data. A residence with a missing QM value was excluded from that respective analysis.
To assess the relationships among the different QMs, a Pearson correlation coefficient r was computed for each unique pair of QMs (Figure). All QMs studied were found to be very weakly or weakly correlated with one another using the Evans classification for very weak and weak correlations (r < 0.20 and 0.20 < r < 0.39, respectively).18
Independent Variables
A total of 15,420 residences were included in the study. Seventy-nine residences did not report occupancy data, however, so those residences were excluded from the occupancy analyses. We categorized the ownership of each nursing home as for-profit, nonprofit, or public. We categorized nursing home size, based on quartiles of the size distribution, as large (> 127 beds), medium (64 to 126 beds), and small (< 64 beds). This method for categorizing the residential characteristics was similar to that used in previous work.19 Similarly, we categorized nursing home occupancy as high (> 92% occupancy), medium (73% to 91% occupancy), and low (< 73% occupancy) based on quartiles of the occupancy distribution. For the regional analysis, we grouped states together based on the CMS regional offices: Atlanta, Georgia; Boston, Massachusetts; Chicago, Illinois; Dallas, Texas; Denver, Colorado; Kansas City, Missouri; New York, New York; Philadelphia, Pennsylvania; San Francisco, California; and Seattle, Washington.20
Analyses
We used Levene’s test to determine whether variances among the residential groups were equal for each QM, using an a priori α = 0.05. For all 20 tests conducted (4 residential characteristics for all 5 QMs), the resulting F-statistics were significant, indicating that the assumption of homogeneity of variance was not met.
We therefore used Welch’s analysis of variance (ANOVA) to evaluate whether the groups within each residential characteristic were the same on their QM means. For example, we tested whether for-profit, nonprofit, and public residences had significantly different mean depression rates. For statistically significant differences, a Games-Howell post-hoc test was conducted to test the difference between all unique pairwise comparisons. An a priori α = 0.05 was used for both Welch’s ANOVA and post-hoc testing. All analyses were conducted in RStudio Version 1.2.5033 (Posit Software, PBC).