Statistical Analysis

The data were analyzed from October 24, 2018, to October 3, 2019. Descriptive analyses were performed for resident characteristics and outcomes using means with SDs for continuous variables and proportions for categorical variables. Descriptive results were generated for each advanced illness group stratified by cohort year. Hospital transfers were further described by type (hospitalizations, observation stays, and ED visits) and by condition (all causes, potentially avoidable, and serious bone fractures). The proportion of potentially avoidable transfers attributable to specific diagnoses (eg, sepsis, pneumonia) were also calculated and presented graphically. Outliers defined as residents with all-cause hospital transfer rates exceeding 365 transfers per person-year alive across 12 months were removed (<1% of residents across all cohorts combined). Hospital transfer outcomes across 12 months were measured as the proportion of residents who experienced at least 1 transfer and the number of transfers per person-year alive. Mortality and hospice enrollment outcomes across 12 months were measured as the proportion of residents who experienced the event and the time to event. For all models, the main independent variable was the cohort year, with 2011 as the referent category. All models were adjusted for age, sex, nonwhite race, and MRS3 score. All models were fitted using generalized estimating equations to account for clustering within nursing homes and included an offset for log-transformed person-time.

Binary outcomes (any hospital transfers, hospice enrollment, and mortality) were analyzed using log-linked binomial models to estimate relative risk with cohort year. Zero-inflated Poisson regression models were used to analyze outcomes measured as number of transfers per person-year alive to allow for overdispersion owing to the high proportion of residents without a hospital transfer.³⁰ Adjusted risk ratios (aRRs) and 95% CIs were generated for these analyses. Finally, Cox proportional hazards regression models were used to analyze time-to-event outcomes (hospice enrollment, mortality) across 12 months of follow-up. In the hospice model, death without hospice was considered to be a competing risk. Adjusted hazard ratios and 95% CIs were estimated from these analyses. Analyses were performed using SAS, version 9.4 (SAS Institute Inc).