Evaluation of US Hospital Episode Spending for Acute Inpatient Conditions After the Patient Protection and Affordable Care Act

Key Points Question Were policy reforms enacted as part of the US Patient Protection and Affordable Care Act (ACA) and budget sequestration associated with reductions in hospital episode spending for acute inpatient hospitalizations? Findings In this policy evaluation that included 7 634 242 episodes from Medicare beneficiaries, alternative estimates indicated that reforms following the ACA were associated with changes in spending between −$431 and −$1232 per episode (approximately −3% to −10%). Cuts to Medicare reimbursement accounted for most of this reduction. Meaning These findings suggest that reforms following the ACA were associated with large reductions in Medicare spending.


Introduction
The US Patient Protection and Affordable Care Act (ACA)-signed into law in March 2010-included numerous provisions to improve hospital quality and reduce Medicare spending for acute episodes.
Many of these reforms, including Hospital Value-Based Purchasing, 1

the Hospital Readmissions
Reduction Program (HRRP), 2 and the Bundled Payment for Care Improvement, 3 disproportionately targeted 3 medical diagnoses: acute myocardial infarction, heart failure, and pneumonia. 4 Other provisions, such as mandated payment cuts, applied more broadly across hospital care. 5,6 In 2013, additional cuts to Medicare fees through the budget sequestration process took effect. 7 The association between the ACA, budget sequestration, and Medicare spending for acute episodes is unclear. Increases in care coordination and efficiency during the episode, potentially spurred by payment reforms like the HRRP, the Medicare Shared Savings Program, and bundled payment programs, may have reduced hospital spending. 2,[8][9][10] At the same time, the effects of the HRRP may be overstated through high comorbidity coding 11 and secular changes in admission rates 12 ; the effectiveness of bundled payment outside of joint replacement appears to be limited 13 ; and the Medicare Shared Savings Program has not effectively engaged hospitals 14 and may be biased by selective attrition. 10 Medicare fee cuts following these reforms reduced spending for specific services but may be offset by the use of additional services elsewhere in the episode. [15][16][17] As a result, the net association of reforms following the ACA with hospital episode spending is uncertain.
In this context, we used national Medicare data to evaluate the association between reforms following the passage of the ACA and hospital episode spending. We used 3 alternative estimation strategies to test for changes in hospital episode spending after the initiation of the ACA.

Data Sources
We used US national Medicare claims from a random 20% sample of Medicare beneficiaries, including data from the Carrier, Outpatient, and Medicare Provider Analysis and Review research identifiable files. We included claims from fee-for-service beneficiaries discharged between January 1, 2008, and August 31, 2015. Additionally, we merged these data with the Annual Survey of the American Hospital Association 18

Study Cohort
Our study cohort included Medicare fee-for-service beneficiaries aged 66 years or older who were enrolled in both Medicare Part A and Part B and had discharges between January 1, 2008, and August 31, 2015. We defined index hospitalizations and readmissions using the specifications similar to the CMS 30-day readmission measures. Specifically, a patient's first admission during the study period is an index admission. Any admission that occurs within 30 days of the index admission is a readmission. Consequently, admissions were classified as either an index admission or a readmission but not both. Patients who were transferred to another hospital, died during the episode, were discharged against medical advice, enrolled in Medicare Advantage, were enrolled in both Medicare and Medicaid (because of potential omissions in episode spending due to payment from Medicaid), did not have at least 12 months of claims data prior to their admission (necessary for severity adjustment), or did not have 30 days of postdischarge data available were excluded.
Targeted medical diagnoses (ie, acute myocardial infarction, heart failure, and pneumonia) were identified using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes that were applied in prior studies. 2,20 Discharges after August 31, 2015, were excluded to avoid changes associated with the retirement of the ICD-9 and the implementation of

International Statistical Classification of Diseases and Related Health Problems, Tenth Revision
(ICD-10) (eFigure 1 in the Supplement).

Measuring Episode Spending
Our primary outcome was price-standardized 30-day total episode spending. This was defined as the sum of payments related to the index hospitalization (diagnosis-related group [DRG] payments and outlier payments), physician services (inpatient and outpatient), readmissions, hospital outpatient care, and postacute care from the date of discharge through 30 days following discharge. We also examined each of these spending components as a separate outcome.
All payments were price-standardized by removing adjustments to payments that were associated with Indirect Medical Education, the Disproportionate Share Index, and with geographybased cost of living and wage indices. Price standardization was performed using methods originally described by the Medicare Payment Advisory Commission and subsequently by the Dartmouth Institute. [21][22][23] This approach has been applied in several prior studies assessing Medicare payments. 24,25 To account for inflation across the 7 years of payment data included in this study, all amounts were adjusted to 2015 US dollars.

Statistical Analysis
Our approach to estimating the association between reforms following the ACA and episode spending was motivated by an attempt to identify the most appropriate counterfactual outcome. In other words, we sought to identify what hospital spending would have been had the ACA and subsequent reforms not been implemented. We used 3 different approaches to estimate this association using different counterfactuals.
Acute care hospitals, paid under the Inpatient Prospective Payment System, were subject to numerous payment reforms under the ACA. Within the population of acute care hospitals, we tested whether changes in spending were greater for commonly targeted diagnoses compared with untargeted diagnoses. This analysis excluded discharges for chronic obstructive pulmonary disease and lower extremity joint replacement. These diagnoses were included in the HRRP beginning in fiscal year 2015 and, therefore, did not fit clearly into the targeted and untargeted classification. By examining changes in spending among only acute care hospitals, this analysis addresses challenges associated with comparing spending across very different types of hospitals. At the same time, estimates from this specification could underestimate the changes associated with reforms following the ACA given that some provisions, such as Medicare fee cuts, apply to the broad spectrum of hospital care rather than the 3 common diagnoses that have been commonly been targeted in ACA programs. In addition, spillovers that reduced spending for untargeted diagnoses due to broad care improvements would also bias estimates from this approach. For this reason, estimates from this specification represent a lower bound of the association of the ACA with episode spending.
Second, we tested whether changes in spending were different between acute care and critical access hospitals. Critical access hospitals are small, rural hospitals that receive cost-based reimbursement (rather than prospective payment). These hospitals were exempt from nearly all ACA provisions targeting hospital spending. While critical access hospitals care for patients with many common medical conditions, including diagnoses targeted under the ACA, they differ from acute care hospitals across a range of characteristics, including size, rurality, and teaching status. The patients seen in critical access hospitals tend to have lower severity, and critical access hospitals tend to provide less intensive and technologically sophisticated care. However, recent work has shown that critical access hospitals perform similarly to acute care hospitals on important measures, including mortality rates for common surgical procedures. 26 In addition, differences between acute care and critical access hospitals remained similar during the study period, limiting the bias associated with comparing outcomes across different populations.
For these 2 comparisons (targeted diagnoses vs untargeted diagnoses and acute care hospitals vs critical access hospitals), we performed difference-in-differences (DID) analyses, estimating linear regression models at the patient-episode level. We estimated separate models for total spending and each component of spending. Our models controlled for age, sex, race, season of admission, severity measured by the hierarchical condition category risk score, diagnosis (based on Agency for Healthcare Research and Quality clinical classification software), 27 hospital characteristics (ie, size, teaching status, Medicaid share of patients, profit status), and hospital fixed effects (eAppendix in the Supplement). The DID models comparing acute care with critical access hospitals controlled for DRG weights instead of clinical classification. We considered the beginning of the reforms to be April 1, 2010, the month following the passage of the ACA. Budget sequestration began on March 1, 2013.
We did not independently assess the association of budget sequestration with episode spending. A timeline of the relevant provisions occurring during the study period is included in eFigure 2 in the Supplement.
To evaluate the validity of our DID approach, we tested whether trends in the study outcomes were parallel for both DID analyses. 28 When comparing spending for targeted and untargeted outcomes, trends were parallel with the exception of inpatient physician services (eTable 1 in the Supplement). When comparing spending between acute care and critical access hospitals, trends were also similar but statistically different for total episode spending and several components of episode spending (eTable 1 and eFigures 3-8 in the Supplement).
To address concerns that the challenges associated with the 2 DID approaches (including nonparallel trends) may lead to bias, we performed a third analysis using a generalized synthetic control approach. 29 Originally developed for a single unit subject to treatment, the generalized synthetic control approach allows for multiple treated units (hospitals), as we have in our study. 29 This approach was developed to identify a comparison group to minimize bias, particularly in cases when parallel trends were violated and the researcher had numerous untreated units that could potentially serve as controls. 24 Under this approach, an interactive fixed-effect model is estimated by using control group data. Then, estimates of factor loadings for each treated unit are used to impute the counterfactual for the treatment group (ie, the value of the outcome for the treatment group in the postperiod if they did not receive treatment). Because panel data are required for synthetic control methods, we performed the generalized synthetic control analysis at the hospital-quarter level. We used mean age, sex, race, hierarchical condition category score, DRG weights, profit status of hospitals, and proportion of Medicare and Medicaid days as covariates in the model (eAppendix and eFigures 9-11 in the Supplement). We then specified the generalized synthetic control model to be created on the basis of episode spending in the 9 quarters prior to the start of the ACA.
We performed a supplemental analysis to approximate the contribution of Medicare's fee reductions to hospital episode spending. Our analysis of changes in spending related to the index hospitalization is an approximation of Medicare fee reductions. However, this approximation may not be entirely accurate given that the index hospitalization includes outlier payments and could be affected by changes in the composition of DRGs over time. To address this, we created a hospital inpatient price index consisting of the 20 most common DRGs during the study period. We then performed a DID analysis, comparing changes in the price index between acute care and critical access hospitals (eAppendix and eTable 2 in the Supplement).
To test the robustness of our findings, we estimated models without accounting for coded severity of illness, which has been shown to inflate the association between other ACA reforms and outcomes. 12 We also estimated models that did not standardize prices across facilities, and to reduce the influence of outliers, Winsorized spending at the 5th and 95th percentiles. Finally, because many of the reforms in the ACA focused on traditional Medicare (rather than Medicare Advantage), the association of the ACA with outcomes may have been stronger in areas with more traditional Medicare enrollment. To explore this, we estimated the association between reforms following the ACA and spending among hospitals in Hospital Referral Regions that had the highest and lowest Medicare Advantage penetration. We used our estimates to approximate the annual savings to Medicare that resulted from reforms following the ACA. To do this, we multiplied our effect estimates by the number of annual episodes to which the estimate applied in the postreform period.

JAMA Network Open | Health Policy
All reported P values were 2-sided, and P < .05 was used as threshold for significance. DID analyses were performed with Stata version 15 (StataCorp). The generalized synthetic control models were estimated using the gsynth package in R version 3.6.3 (R Project for Statistical Computing). 30 Analysis of this data was initiated February 1, 2019 and completed on July 8, 2020.

Results
A total of 7 634 242 episodes (4 525    Results from models examining the association between reforms following the ACA and components of episode spending are shown in Table 3. Estimates show that the largest reductions were related to the index admission, followed by spending related to readmissions and postacute care. Supplemental analysis found that spending reductions related to the index admission were largely associated with cuts to Medicare fees (eTable 4 in the Supplement).  Abbreviations: DID, difference-in-differences. a 95% CIs for DID models are based on standard errors that are robust to clustering at hospital level. Standard errors for generalized synthetic control models were estimated using bootstrapping with 1000 iterations. Targeted diagnoses include acute myocardial infarction, heart failure, and pneumonia. Discharges for chronic obstructive pulmonary disease and lower extremity joint replacement were excluded from acute care only DID models. Models estimated among targeted diagnoses controlled for whether targeted condition was acute myocardial infarction, heart failure, or pneumonia.   Abbreviations: DID, difference-in-differences; NA, not available. a 95% CIs for DID models are based on standard errors that are robust to clustering at hospital level. Standard errors for generalized synthetic control models were estimated using bootstrapping with 1000 iterations. NA indicates that the estimate was not available due to failure of model to converge (eAppendix in the Supplement). Targeted diagnoses include acute myocardial infarction, heart failure, and pneumonia. Discharges for chronic obstructive pulmonary disease and lower extremity joint replacement were excluded from acute care only DID models. Models estimated among targeted diagnoses controlled for whether targeted condition was acute myocardial infarction, heart failure, or pneumonia.

Discussion
Our study has 3 principal findings that improve our understanding of reforms following the ACA.
First, across several alternative estimation strategies, reforms following the ACA were associated with substantial reductions in hospital episode spending. Second, reductions in episode spending were the greatest for the index admission and readmission hospitalization spending. The substantial reductions in spending for the index admission coupled with other evidence suggests that Medicare fee reductions were associated with most of the reductions in episode spending. Third, spending reductions were substantially greater in areas with lower Medicare Advantage penetration.
To our knowledge, ours is the first study to evaluate whether reforms following the ACA were associated with reductions in hospital episode spending. Our findings are consistent with national data showing virtually no growth in hospital inpatient spending following the ACA 31 and greater spending growth for critical access hospitals compared with acute care hospitals. 32  Reductions in readmission rates and in the use of postacute care after discharge likely also contributed to savings. However, although we did not estimate this directly, Medicare fee reductions would also contribute to the associations of reforms with spending related to readmissions. As a result, we conclude that reforms following the ACA were associated with reductions in hospital episode spending, primarily through Medicare fee reductions for hospitals and through reductions in spending for postacute care.

Limitations
Our study should be interpreted in the context of several important limitations. First, we were not able to evaluate the association of reforms following the ACA with spending using an ideal control group, ie, hospitals that were similar to acute care hospitals but not exposed to reforms. Instead, our estimates used a combination of untargeted diagnoses and critical access hospitals (not exposed to ACA reforms) as the counterfactual. Nonetheless, across a wide range of estimation strategies, we

JAMA Network Open | Health Policy
Hospital Episode Spending for Acute Inpatient Conditions After the ACA found consistent evidence that reforms following the ACA were associated with substantial reductions in spending.
Second, by using administrative data, we cannot fully account for patient risk, which can drive spending patterns. However, using administrative data sources allows us to closely adhere to the same data used to administer and set penalties under Medicare reforms. 34 A sensitivity analysis found that our results were not sensitive to comorbidity coding. Third, because we did not have access to the data, our study did not evaluate Medicare spending associated with home health care and durable medical equipment. 35 Because the CBO estimated that the ACA led to $0.75 billion in annual reductions related to home health spending during our study period, our overall estimates may be understated. 17 In addition, our analysis of total spending reductions assumed that the reforms we considered had no effect on the volume of admissions in the postreform period.
Quarterly volume declined for acute care hospitals from approximately 253 000 in the prereform period to 235 000 in the postreform period. ACA reforms focusing on efficiency could have contributed to these reductions. At the same time, hospitals may have tried to increase volume to counteract fee cuts. Because the magnitude of these countervailing effects are ambiguous, we did not consider them in our estimates of total spending reductions.
Fourth, we examined hospital spending within 30 days of discharge, rather than the 90-day period now commonly used in bundled payment programs. However, 30-day episode spending is commonly used in national programs, such as Hospital Compare 36 and Hospital Value-Based Purchasing. In addition, between 75% and 84% of 90-day episode spending occurs within a 30-day episode. 37 For these reasons, 30-day episodes are appropriate to assess hospital spending, and the use of longer episode durations are unlikely to affect our conclusions. Medicare fee reductions for physician, hospital outpatient, and postacute care services also contributed to reductions in episode spending, although we did not estimate the precise magnitude of these fee reductions. Seventh, given the many myriad changes that occurred in US health care around the time of the ACA, we are limited in our ability to determine the precise mechanism that led to reductions in Medicare spending.

Conclusions
This policy evaluation has important implications for stakeholders seeking to understand the impact of policies to reduce spending in Medicare. This study found that a combination of provisions in the ACA and budget sequestration, likely associated with overall payment reductions and penalties as well as reforms focused on quality improvement, were associated with reductions in Medicare spending. This suggests that multipronged policy changes focused on conditional or unconditional payment reductions can generate substantial savings.