Medicare's hospital readmissions reduction program and the rise in observation stays

Abstract Objective To evaluate whether Medicare's Hospital Readmissions Reduction Program (HRRP) is associated with increased observation stay use. Data Sources and Study Setting A nationally representative sample of fee‐for‐service Medicare claims, January 2009–September 2016. Study Design Using a difference‐in‐difference (DID) design, we modeled changes in observation stays as a proportion of total hospitalizations, separately comparing the initial (acute myocardial infarction, pneumonia, heart failure) and subsequent (chronic obstructive pulmonary disease) target conditions with a control group of nontarget conditions. Each model used 3 time periods: baseline (15 months before program announcement), an intervening period between announcement and implementation, and a 2‐year post‐implementation period, with specific dates defined by HRRP policies. Data Collection/Extraction Methods We derived a 20% random sample of all hospitalizations for beneficiaries continuously enrolled for 12 months before hospitalization (N = 7,162,189). Principal Findings Observation stays increased similarly for the initial HRRP target and nontarget conditions in the intervening period (0.01% points per month [95% CI −0.01, 0.3]). Post‐implementation, observation stays increased significantly more for target versus nontarget conditions, but the difference is quite small (0.02% points per month [95% CI 0.002, 0.04]). Results for the COPD analysis were statistically insignificant in both policy periods. Conclusions The increase in observation stays is likely due to other factors, including audit activity and clinical advances.

What is known on this topic • Observation stays have increased significantly over the last 15 years.
• Observation stays are excluded from the numerator and denominator of readmissions measures.
• Medicare's Hospital Readmissions Reduction Program (HRRP) may incentivize hospitals to substitute observation stays for inpatient readmissions, although evidence of this is mixed.

What this study adds
• Ours is the first study to examine how hospitals may have responded to the HRRP by using observation stays in lieu of both index admissions and readmissions.
• Using quasi-experimental methods, we find that the HRRP was not meaningfully associated with differential increases in the use of observation stays for target versus nontarget conditions.
• The increase in observation stays is more likely the result of advances in clinical care and hospitals taking steps to avoid payment audits.

| INTRODUCTION
Medicare's Hospital Readmissions Reduction Program (HRRP) has been controversial since its announcement. One ongoing concern is whether the HRRP incentivizes hospitals to use observation stays to avoid readmission penalties, as observation stays are excluded from the numerator and denominator of readmission measures. 1 Approximately 23% of 30-day return hospitalizations-and 18% of those with HRRP target conditions-are not captured in readmissions performance because they occurred after discharge from an observation stay or a patient's return hospitalization was classified as observation. 2 Only some HRRP studies have found 30-day readmission and observation rates to be negatively correlated. [3][4][5] One regression discontinuity evaluation found nonpenalized hospitals used significantly more observation stays for heart failure patients just before versus after the 30-day HRRP cutoff. 6 Observation stays at nonpenalized hospitals were longer, suggesting potential substitution of observation stays for otherwise longer or more complex inpatient readmissions. 7 Importantly, all prior evaluations only considered observation stays in the 30 days following discharge from an inpatient admission. To date, no studies have examined how hospitals may have responded to the HRRP by using observation stays in lieu of both index admissions and readmissions.
In this study, we used a difference-in-difference (DID) approach to examine whether HRRP implementation was associated with an increased rate of observation stay use among Medicare beneficiaries.

| METHODS
Using a 20% sample of fee-for-service Medicare claims between January 1, 2009 and December 31, 2016, we identified inpatient admissions and hospital observation stays to short-term general and specialty hospitals for beneficiaries continuously enrolled for 12 months prior to hospitalization. Although these data are more than 5 years old, we use them because they include the periods immediately before and after HRRP announcement and implementation. We derived our sample by adapting the inclusion and exclusion criteria of CMS' hospital-wide readmission measure. 8 We excluded all hospitalizations to cancer specialty hospitals, critical access hospitals, and Maryland hospitals (both of which are exempt from the HRRP), and hospitals with fewer than 100 inpatient admissions in the prior year (to avoid unstable estimates). We also excluded 96 outlier hospitals with abnormally high rates of observation (defined as those with rates of observation stay use at least 1.5 times the interquartile range for the national sample). Among the remaining hospitals, we excluded individual hospitalizations for hip and knee replacement (which were added to the HRRP later), a primary diagnosis including psychiatric conditions, rehabilitation, or cancers, and cases in which the beneficiary transferred to another facility, left the hospital against medical advice, or died in the hospital (N = 2,809,281). Finally, given evidence that CMS coding changes during HRRP implementation (which increased the number of diagnoses fields on a claims record from 9 to 25) 9 may overestimate the HRRP effect, we limited capture of comorbidities to the first 9 diagnosis codes.
We identified hospitalizations for HRRP target conditions using clinical classification software (CCS) codes. 10 We combined the first 3 conditions initially targeted by the HRRP (acute myocardial infarction, pneumonia, heart failure) into a single group for analysis. We analyzed COPD separately as it was added to the program later. To select an appropriate control group from conditions not targeted by the HRRP, we examined observation stays (defined by the presence of revenue center code 0760 or 0762) as a proportion of total hospitalizations for each condition in the pre-period. Hospitalizations were classified as observation stays or inpatient admissions based on their status at discharge (i.e., observation stays that converted to inpatient admissions were classified as inpatient admissions). We observed substantial heterogeneity in this measure across conditions ranging from near 0% to over 50% of all hospital events. HRRP target conditions are predominantly managed via inpatient admission and had relatively low rates of observation use. Conditions with the highest rates of observation tended to be symptom diagnoses (e.g., chest pain, dizziness, syncope), which are qualitatively different from the targeted conditions. Thus, we elected to use a control group consisting of nontargeted conditions with observation stays comprising less than 10% of total hospitalizations in the pre-period as our primary approach.
This allowed us to compare changes in target and nontarget conditions primarily managed as inpatient admissions. However, we recognize this threshold is arbitrary, and conducted a sensitivity analysis using an alternative control group of all nontargeted conditions with parallel pre-period trends. As shown in equation 1, we estimated linear probability models within a DID framework to examine if the HRRP increased observation stay use among targeted conditions:

| RESULTS
Our primary analysis included 7,162,189 hospitalizations, of which 18.4% were for one of the first 3 HRRP target conditions. Unadjusted trends in observation stay use among our two target condition groups and our nontarget control group are shown in Figure 1. Table 1  The results of our fully adjusted DID models appear in Table 2    hospitalization may make programs such as the HRRP less representative of hospital quality and use patterns over time and may misrepresent true performance based on local observation practices. 1