Assessment of Hospital Readmissions From the Emergency Department After Implementation of Medicare’s Hospital Readmissions Reduction Program

Key Points Question Is Medicare’s Hospital Readmissions Reduction Program associated with changes in the probability of readmission for discharged patients revisiting the emergency department (ED)? Findings In this cohort study including data for 9 914 068 patients, the Hospital Readmissions Reduction Program was associated with a 1.4% decrease in readmissions at the ED revisit. The program was associated with 1.2% fewer readmissions from the ED involving clinical presentations for which admission is usually indicated. Meaning These findings suggest that the Hospital Readmissions Reduction Program was associated with changes in the probability of readmission for certain conditions, highlighting the critical and evolving role of the ED in hospital readmission patterns.


eAppendix 1. Parallel Slopes Assumption
In order for difference-in-differences to work properly, we need to satisfy the parallel slopes assumption. Specifically, this states that during the baseline period, readmissions from the ED of recently-discharged patients should be similar for the two comparison groups (the "intervention" group, ie patients at the index hospitalization who have conditions targeted by HRRP, versus the "control" group, ie patients at the index hospitalization who have conditions that are not targeted by HRRP).
Below are the unadjusted trends in admission before HRRP implementation. These suggest that the parallel slopes assumption was met.
In addition, we empirically tested the parallel slopes assumption by narrowing the sample to the pre-period (ie limiting the sample to before HRRP implementation) and falsely creating a "post" indicator for January 1, 2011. The pre-period placebo test yielded a difference-in-difference estimate of -0.001 (95% CI: -0.007, 0.004, p=0.621), which suggests that the groups did not change in trends before and after the false post indicator of January where Y is whether the patient was readmitted during the ED revisit; Post is whether the ED revisit was after HRRP implementation; Target is whether the patient's initial hospitalization was for a condition targeted by HRRP; Risk are patient risk adjustment factors, including: age, sex, race, Charlson comorbidity index, the readmission cohort (as defined by CMS), and whether the ED revisit was on a weekend; Hosp are hospital characteristics, including academic status, size, urbanicity, and ownership; State is the state in which the hospital is located; and Year is the year the discharge patient revisited the ED. Age was by 5-year bands until age 85; all patients 85 years or older were grouped together. Academic status was either an accredited program in the Accreditation Council for Graduate Medical Education or Member of Council of Teaching Hospital of the Association of American Medical Colleges. Hospital size was based on quintiles of admission. Metropolitan or urban designation was based on the Rural-Urban Continuum Codes, which we further classified into whether the hospital was in a metropolitan or urban nonmetropolitan area (as described in the paper, we excluded rural hospitals).

eAppendix 3. Robustness Checks
We conducted several robustness checks. First, we examined alternative model specifications (logistic model and hospital fixed effects). Second, we modified the outcome to account for potential differences in care patterns. Specifically, we defined the outcome variable to be admission or transfer, or admission or observation stay. Observation stay information was available only for two of the three states in our study (Florida and New York). Third, previous research 1-3 suggests that community factors, including availability of post-acute care, influence readmissions. Following previous research, we included additional covariates in our model for population size, unemployment, median household income, percent poverty, percent uninsured, percent of the population in deep poverty, population in nursing homes, percent non-English speaking population, percent vacant housing, educational attainment, number of nursing facilities per 100K, primary care providers per 100K, number of specialists per 100K, number of cardiologists per 100K, number of community health centers per 100K, the ratio of primary care physicians to specialists, the Medicare population per 100K, the number of federally-qualified health centers per 100K, and the number of hospital beds per 100K. Fourth, in the main analysis, we adjusted for changes in electronic transmission standards that increased the number of diagnosis codes and that may make patients appear more sick. 4 We did this by adjusting our Charlson index so that they use only the first 9 secondary diagnosis codes. In the robustness check, we used all secondary diagnosis codes in calculating the Charlson index. Fifth, we added dummy variables for time (represented by separate variables for each month of the study period, except the first) to account for overall time trends. Note that we do not include this in the main analysis because only two states provided month. Sixth, we examined our exclusion of conditions later targeted by HRRP. Seventh, we conducted a falsification test where we examine readmission patterns for patients revisiting the ED 31 to 45 days after admission. Eighth, for conditions at the ED where admission is more variable, we exclude mood disorders.
The results from these robustness checks are below.   Notes. Linear probability model with robust clustered standard errors for the hospital. The model controls for age, sex, race/ethnicity, Charlson comorbidity index, whether the ED revisit was on a weekend, patient's condition at the index hospitalization, hospital size, teaching status, metropolitan or urban designation, ownership status, state, and year. In addition, each of the sensitivity analyses includes the variables as described above. A negative difference-in-difference indicates that readmissions from the ED decreased post-implementation for recently-discharged patients. a Discretionary ED diagnoses are five ED diagnoses for which patient admissions to the hospital are more variable: mood disorders, nonspecific chest pain, skin and soft tissue infections, urinary tract infections, and COPD.