University of Southern California Dissertations and Theses

Health care utilization and spending of the U.S. aging population

(USC Thesis Other)

Health care utilization and spending of the U.S. aging population

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Health Care Utilization and Spending of the U.S. Aging Population
by
Shengjia Xu
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(PUBLIC POLICY AND MANAGEMENT)
May 2024
Copyright 2024 Shengjia Xu

ii
ACKNOWLEDGEMENTS
I am extremely grateful to my dissertation committee chair, Alice Chen, who guided me
through this long journey step by step with her unwavering support. She is the best mentor I
could have: always available and offers invaluable advice on conducting and presenting my
research as well as collaborating with others.
I am also deeply indebted to my other committee members, Julie Zissimopoulos and
Darius Lakdawalla. Their expertise, constructive criticism, and insightful feedback constantly
improved my work. I also benefited tremendously from working with Julie on other projects
which expanded my research interests and connected me to people and resources.
I also have the fortune to work with and learn from a few other faculty members on
different projects during my doctoral journey, including Mireille Jacobson, Jakub Hlavka, Alex
Capron, and Michelle Keller. Their mentorship and advice have made me a better researcher. My
appreciation also goes to Jillian Wallis, who kindly aided in my access to restricted data and
helped me with dissertation data reuse applications. I also received a lot of help from Julie Kim
and Anna Parks, who helped me navigate many administrative issues.
I would like to thank my colleagues and friends, Yimin Ge, Yi Chen, Niloofar FouladiNashta, Johanna Thunell, Patricia Ferido, Sidra Haye, Haley Garland, Drishti Baid, Jianhui Xu,
and Alison Holt for their camaraderie and support.
This journey wouldn’t be possible without the love and encouragement from my family. I
want to thank my parents, who have always been supportive of every decision I made for my
education and career. I also want to thank my parents-in-law for their patience, kindness, and
many home-made meals. Finally, I want to thank my husband, Elbert Pu, for his love, support,
and hand poured coffees that kept me through long days of research and writing.

iii
Table of Contents
ACKNOWLEDGEMENTS.......................................................................................................... ii
LIST OF TABLES........................................................................................................................ iv
LIST OF FIGURES ....................................................................................................................... v
Chapter 1: Introduction.................................................................................................................. 1
Chapter 2: Do ACOs Better Manage Comorbid Chronic Conditions of Patients with Serious
Mental Illnesses? Evidence from the Medicare Shared Savings Program .................................... 6
INTRODUCTION ..................................................................................................................... 6
METHODS ................................................................................................................................ 9
RESULTS ................................................................................................................................ 16
DISCUSSION.......................................................................................................................... 23
TABLES & FIGURES............................................................................................................. 27
APPENDIX.............................................................................................................................. 35
Chapter 3: Impact of Medicare Shared Savings Program on Health Care Spending and
Utilization Using Regression Discontinuity Design.................................................................... 49
INTRODUCTION ................................................................................................................... 49
METHODS .............................................................................................................................. 51
RESULTS ................................................................................................................................ 57
DISCUSSION.......................................................................................................................... 61
TABLES & FIGURES............................................................................................................. 65
APPENDIX.............................................................................................................................. 70
Chapter 4: Association of Dementia Severity at Diagnosis with Health Care Utilization and Costs
around the Time of Incident Diagnosis........................................................................................ 90
INTRODUCTION ................................................................................................................... 90
METHODS .............................................................................................................................. 92
RESULTS ................................................................................................................................ 98
DISCUSSION........................................................................................................................ 103
TABLES & FIGURES........................................................................................................... 108
APPENDIX............................................................................................................................ 114
REFERENCES .......................................................................................................................... 133

iv
LIST OF TABLES
Table 1. Beneficiary Characteristics by ACO Alignment and SMI Status, 2008-2011 .............. 28
Table 2. Effect of ACOs on Health Care Utilization................................................................... 29
Table 3. Effect of ACOs on Health Care Utilization by Comorbid Chronic Conditions ............ 30
Table 4. Effect of ACOs on Health Care Spending..................................................................... 32
Table 5. Effect of ACOs on Health Care Spending by Comorbid Chronic Conditions .............. 33
Table 6. Sample Characteristics at Baseline in Relation to the 2013 ACO Retrospective
Alignment Threshold ................................................................................................................... 66
Table 7. Discontinuities in Health Care Spending in relation to 2013 and 2012 ACO
Retrospective Alignment Thresholds........................................................................................... 68
Table 8. Discontinuities in Health Care Utilization in relation to 2013 and 2012 ACO
Retrospective Alignment Thresholds........................................................................................... 69
Table 9. Sample Selection Criteria for the Study Population .................................................... 108
Table 10. Characteristics of Study Population by Dementia Severity at Diagnosis.................. 109

v
LIST OF FIGURES
Figure 1. Sample Composition by ACO Alignment Length, 2012-2015 .................................... 27
Figure 2. Sample Composition by ACO Alignment Years, 2012-2015 ...................................... 27
Figure 3. Effect of ACOs on Health Care Utilization.................................................................. 31
Figure 4. Effect of ACOs on Health Care Spending.................................................................... 34
Figure 5. Probability of ACO Alignment as a Function of the Alignment Threshold................. 65
Figure 6. MSSP ACO-Related Discontinuities in Annual Health Care Spending, 2013 ACO
Retrospective Alignment Threshold ............................................................................................ 67
Figure 7. Predicted Outpatient Visits and Costs Before and After Dementia Diagnosis by
Dementia Severity at Diagnosis................................................................................................. 110
Figure 8. Predicted Inpatient Care Utilization and Costs Before and After Dementia Diagnosis by
Dementia Severity at Diagnosis................................................................................................. 111
Figure 9. Predicted Emergency Room Utilization Before and After Dementia Diagnosis by
Dementia Severity at Diagnosis................................................................................................. 113

1
Chapter 1: Introduction
My dissertation explores two topics in health care of the U.S. aging population, with
health care utilization and spending as the underlying theme. The first topic is the Medicare
Shared Savings Program (MSSP) implemented by the Centers for Medicare and Medicaid
Services (CMS) as part of a broader provider payment reform. The MSSP creates global
incentives for accountable care organizations (ACOs) to reduce health care spending and
improve quality of care for assigned Medicare fee-for-service (FFS) beneficiaries. The MSSP is
becoming increasingly important, as a growing share of Medicare FFS beneficiaries are served
by providers participating in the program. As for 2022, there were 483 ACOs serving 11 million
Medicare FFS beneficiaries (CMS, 2022). My dissertation studies how early cohorts of the
MSSP affected health care utilization and spending of older U.S. adults with serious mental
illnesses (SMI) and other comorbidities. My dissertation also addresses the issue of non-random
ACO participation and identifies the impact of MSSP on health care utilization and spending
among beneficiaries seeing more than one provider for primary care services.
The second topic of my dissertation is dementia, a common age-related chronic condition
that affects 7.6 million older U.S. adults and extracts a heavy financial burden on society
(Alzheimer’s Association, 2023). The association of dementia severity at diagnosis with health
care use and costs around the time of diagnosis is still not well understood, due in part to data
limitations. My dissertation brings together multiple data sources to quantify health care
utilization and spending around the time of dementia diagnosis for older adults diagnosed at
different stages of dementia, thus shedding light on how early diagnosis may impact health care
use and costs for society.

2
In the second chapter, titled “Do ACOs Better Manage Comorbid Chronic Conditions of
Patients with Serious Mental Illnesses? Evidence from the Medicare Shared Savings Program,”
I estimated the impact of early cohorts of MSSP ACOs on health care use and spending among
SMI patients, especially those with other comorbid chronic conditions. The current literature on
ACOs focuses primarily on the entire Medicare FFS population (McWilliams, 2016;
McWilliams, Chernew, et al., 2017; McWilliams, Gilstrap, et al., 2017; McWilliams et al., 2016;
McWilliams et al., 2020; McWilliams et al., 2018). There is little empirical evidence on how
ACOs affect clinically vulnerable populations with multiple comorbidities, such as those with
SMIs who have historically suffered from excessively high health care costs and poor outcomes
(Figueroa et al., 2020), due in part to providers’ lack of financial incentives to coordinate care.
Using difference-in-difference-in-differences and event study approaches, I found that
early cohorts of ACOs generated large and statistically significant reductions in inpatient, skilled
nursing facility (SNF), and emergency room (ER) utilization and associated costs among SMI
beneficiaries who were aligned, even more so than beneficiaries without SMIs. I estimated that
the MSSP resulted in a 5% reduction in inpatient costs, 18% reduction in SNF costs, and 5%
reduction in ER costs among SMI patients. The differential decline was more pronounced among
SMI patients with other comorbid chronic conditions. These results suggest that the MSSP may
have the potential to reduce costs and incentivize provision of coordinated care for people with
SMIs and multiple comorbidities.
The third chapter, titled “Impact of Medicare Shared Savings Program on Health Care
Spending and Utilization Using Regression Discontinuity Design” is coauthored with Alice
Chen, Mireille Jacobson, and Darius Lakdawalla. Prior literature on the impact of early cohorts
of MSSP ACOs all adopted a difference-in-differences framework and found that ACOs

3
achieved very modest reductions in total Medicare spending with no clear evidence on quality of
care for assigned Medicare FFS beneficiaries (McWilliams, 2016; McWilliams et al., 2016;
McWilliams et al., 2020; McWilliams et al., 2018). However, critics have pointed to the
possibility that the formation of ACOs are not random. There is suggestive evidence that
providers with greater capacity, more experience in risk sharing, and serve less disadvantaged
and healthier populations are more likely to participate in the MSSP (Chukmaitov et al., 2019;
Colla, Lewis, Tierney, et al., 2016; Epstein et al., 2014; Yasaitis et al., 2016). Therefore,
identifying the causal impact of ACOs requires an ability to net out differences on observable
and unobservable characteristics that result from selection which previous empirical methods
might fail to do so. Additionally, prior studies included a substantial share of beneficiaries for
whom there may be very little room for ACOs to reduce costs, such as patients who relied on one
provider for primary care services.
We addressed the non-random ACO participation by implementing a regression
discontinuity approach to identify how early cohorts of the MSSP affected health care utilization
and spending among beneficiaries seeing more than one provider for primary care services. We
found that the MSSP reduced total spending by 8% to 12% for these beneficiaries. Health care
utilization also fell considerably, with inpatient stays falling by 13% to 16% and SNF stays
falling by 15% to 20%. Our findings suggest that early cohorts of MSSP ACOs have been
considerably more successful at reducing health care utilization and spending than previously
thought, particularly among patients who stand to gain the most from coordinated care efforts.
The fourth chapter, titled “Association of Dementia Severity at Diagnosis with Health
Care Utilization and Costs around the Time of Incident Diagnosis” is coauthored with Niloofar
Fouladi-Nashta, Yi Chen, and Julie Zissimopoulos. There is considerable evidence that medical

4
costs increase substantially around the time of dementia diagnosis (Lin et al., 2016; White et al.,
2019; C. W. Zhu et al., 2015) which are driven by hospitalizations, ER visits, and post-acute
services (Bynum et al., 2004; Coe et al., 2023; Daras et al., 2017; Desai et al., 2019; Hoffman et
al., 2022; Carolyn W. Zhu et al., 2015; C. W. Zhu et al., 2015). However, these prior estimates
reflect averages of persons in various stages of disease progression, and individuals diagnosed at
different stages of dementia are likely to utilize health care differently. Data limitations have
generally precluded population-based estimates of differences in health care use and costs for
persons diagnosed at different stages of dementia.
Using Health and Retirement Study data linked with traditioanl Medicare claims, we
examined how heterogeneity in dementia severity at diagnosis is related to differences in health
care utilization and spending over a two-year period around the time of incident dementia
diagnosis. With an event study design, we found that acute and outpatient care use and
associated costs were higher for persons diagnosed at mild dementia compared to moderate or
severe dementia before a dementia diagnosis. While higher outpatient utilizaiton and costs
sustained for persons diagnosed at mild dementia in the quarter of diagnosis and afterwards, the
pattern reversed for acute care utilization: persons diagnosed at moderate and severe stages of
dementia were more likely to be hospitalized and incur ER visits relative to mild stage. Acute
care spending, however, was similar for persons diagnosed at different dementia stages
throughout. Contrary to previous thoughts, these results suggest some but limited opportunity for
reducing acute care utilization for persons diagnosed at advanced stages of dementia around the
time of diagnosis, with little to no impact on acute care spending. The results also suggest that
earlier diagnosis may increase use and spending on outpatient care.

5
Taken together, findings from this dissertation provide important insights on health care
utilization and spending of the U.S. aging populaiton, with a special focus on how provider
payment reforms and certain chronic conditions affect the levels and patterns of health care use
and costs of older adults. The findings from my dissertation may inform policy interventions that
aim at reducing health care spending and improving value of care for the aging population.

6
Chapter 2: Do ACOs Better Manage Comorbid Chronic Conditions of Patients with
Serious Mental Illnesses? Evidence from the Medicare Shared Savings Program
INTRODUCTION
Individuals with serious mental illnesses (SMIs), such as schizophrenia, bipolar disorder,
and major depressive disorder, are at significantly higher risk of developing multiple comorbid
chronic conditions, due in part to the long-term use of antipsychotics and antidepressants with
adverse cardio-metabolic effects (Chesney et al., 2014; John et al., 2018; Newcomer, 2005;
Newcomer & Hennekens, 2007; Olfson et al., 2015; Osborn et al., 2007). Compared to the
general population, the prevalence of cardiovascular diseases, diabetes, respiratory diseases,
kidney diseases, and dementia is considerably elevated among persons with SMIs (Brown &
Wolf, 2018; Cai & Huang, 2018; Figueroa et al., 2020; Hagi et al., 2021; Mitchell et al., 2013;
Newcomer & Hennekens, 2007; Osborn et al., 2008; Osborn et al., 2007; Stroup et al., 2021).
The elevated risks of comorbid conditions among SMI persons are heightened by mental
illness-related stigma, access barriers to primary care, fragmented system of health care delivery,
and inferior quality of care (Baller et al., 2015; Corrigan et al., 2014; Happell et al., 2016;
Henderson et al., 2014; Jones et al., 2008; Knaak et al., 2017; McGinty et al., 2015). All these
factors contribute to missed opportunities of early diagnosis and treatment of comorbid
conditions, aggravated disease progression, and greater downstream health care utilization and
costs. SMIs, together with comorbid conditions, extract a heavy financial burden on the health
care system. With $201 billion annual health care expenditure, mental illnesses are already one
of the most costly chronic conditions in the U.S. (Roehrig, 2016), however, this number may
underestimate the total spending caused by mental illnesses because expenditures on comorbid

7
conditions are not fully captured. A recent study finds that SMIs are associated with over 33%
increase in medical spending on comorbid conditions, including congestive heart failure,
ischemic heart disease, diabetes, chronic kidney disease, and chronic obstructive pulmonary
disease (Figueroa et al., 2020). Moreover, the majority of the health care expenditures among
SMI persons does not come from treating and managing SMIs but are associated with comorbid
conditions that are not effectively managed in this population. In fact, non-mental health
expenditures of SMI patients are twice as high as the mental health expenditures, driven by
intensive use of care in acute and post-acute settings, including hospitalizations, skilled nursing
facility (SNF) stays, and emergency room (ER) visits (Figueroa et al., 2020).
As one of the nation’s primary alternative payment models, the Medicare Shared Savings
Program (MSSP) implemented by the CMS presents unique opportunities to improve care
coordination and disease management for older adults with SMIs. Unlike the traditional fee-forservice (FFS) payment which offers little financial incentive for providers to coordinate care, the
MSSP is designed to incentivize voluntarily formed accountable care organizations (ACOs) to
deliver effective care coordination and disease management (CMS, 2014). Since under the MSSP
contract, participating ACOs are accountable for the total costs of assigned Medicare FFS
beneficiaries and have the opportunity to earn financial rewards from CMS if they achieve global
cost reductions, they are incentivized to provide effective upstream care with strong a focus on
primary care to prevent use of more expensive acute and post-acute care (McWilliams et al.,
2013; McWilliams, Gilstrap, et al., 2017; Rittenhouse et al., 2009). Given the excess chronic
disease burden and intensive use of acute and post-acute care among SMI persons, it is of ACO’s
financial interest to better manage comorbid chronic conditions for this highest-cost population.

8
To date, however, there is little empirical evidence on how MSSP ACOs managed
comorbid chronic conditions of SMI persons. The literature on ACOs have focused on total
savings and, to some extent, quality of care, achieved across the entire Medicare FFS population
(McWilliams, Chernew, et al., 2017; McWilliams et al., 2016; McWilliams et al., 2020;
McWilliams et al., 2018; Nyweide et al., 2015). There is considerable evidence that MSSP
ACOs resulted in modest reduction in total Medicare spending with no clear evidence of
worsened quality of care. The magnitude of cost reduction varies by MSSP entry years and
ACO’s organizational structure (Colla, Lewis, Kao, et al., 2016; McWilliams et al., 2016;
McWilliams et al., 2020; McWilliams et al., 2018).
Evidence is still sparse for clinically vulnerable populations with multiple comorbidities,
including individuals with SMIs and other chronic conditions, even though they present the
greatest opportunities to generate financial rewards for ACOs. So far, only two studies examined
the effects of ACOs on high-risk beneficiaries and they found contradictory results (Colla,
Lewis, Kao, et al., 2016; McWilliams, Chernew, et al., 2017). Moreover, it remains unclear
whether ACO’s achievements in persons affected only by medical conditions could be replicated
in SMI patients who present more multifaceted clinical profiles and health care needs.
In addition, three studies have examined the impact of ACOs on persons with mental
illnesses. However, these studies focused exclusively on mental health outcomes (Acevedo et al.,
2021; Busch et al., 2016; Busch et al., 2017). There is some evidence that ACOs resulted in
better diagnosis of depression and adherence to antidepressants (Busch et al., 2016; Busch et al.,
2017). Another more recent study finds that while ACOs reduced use of inpatient and outpatient
mental health care, likelihood of receiving adequate depression care also decreased (Acevedo et
al., 2021). However, given the well-documented association between mental illnesses and

9
comorbid diseases and the fact that health care spending for comorbidities outweighs those for
mental illnesses, these studies do not fully reflect how ACO providers influence care
coordination and management for individuals with mental illnesses.
In this study, I examined how ACOs affected health care utilization and spending among
individuals with SMI diagnoses, especially those with both SMIs and other comorbid chronic
conditions. Despite our growing knowledge of ACOs, we know relatively little about how
comorbidities factor in ACO’s ability to affect health care utilization and spending. This study
aims to bridge the gap by examining ACO’s effects on health service use and costs in SMI
populations with common medical comorbidities. Hospitalizations, SNF stays, and ER visits are
informative of ACO’s management of comorbidities among SMI persons since acute and postacute care are generally thought as low-value and potentially preventable through provision of
effective primary care and care coordination (McWilliams et al., 2013; McWilliams, Gilstrap, et
al., 2017; McWilliams et al., 2016; Rittenhouse et al., 2009). Evidence provided by this study is
crucial in advancing our knowledge of whether and the extent to which the ACO model could
serve as an effective venue to deliver coordinated care for people with multiple comorbidities
and complex health care needs. Findings from this study could also inform design of future
alternative payment models to better align providers’ financial incentives to benefit clinically
vulnerable populations.
METHODS
Data and Study Population
This study used Medicare administrative claims and enrollment data for annual random
20% samples of FFS beneficiaries from 2008 to 2015. For each year, the study population

10
included Medicare beneficiaries who were continuously enrolled in FFS Parts A and B during
the year while alive as well as in the previous year, not enrolled in Medicare Advantage, living in
the U.S., and actively accessing primary care services from ACO providers. Following the CMS
MSSP beneficiary assignment rules, eligible beneficiaries were retrospectively attributed each
year to provider organizations (ACO or non-ACO) that delivered the largest share of qualified
evaluation and management (QEM) services by primary care practitioners or specialists with
primary care designation (CMS, 2017). Only Part B physician claims were considered for
beneficiary’s ACO alignment in this study. While services from Federally Qualified Health
Centers, Rural Health Clinics, and Critical Access Hospitals are also possible for beneficiary
alignment, they are excluded in previous ACO studies because they involve different beneficiary
alignment algorithms. A list of QEM services and physician specialty codes used in beneficiary
assignment are included in Supplementary Table 2.1 and Supplementary Table 2.2, respectively.
Instead of defining ACO as a collection of taxpayer identification numbers (TINs), this
study used an alternative ACO definition as a collection of physician National Provider
Identifiers (NPIs). As showed in previous literature, defining ACOs based on TINs or NPIs has
little impact on outcomes (McWilliams et al., 2016; McWilliams et al., 2020; McWilliams et al.,
2018). Using the ACO Provider-level Research Identifiable Files (RIF), I identified physician
NPIs associated with each ACO organization in 2012-2015 and collapsed the QEM expenditures
at beneficiary-provider organization level to determine whether a beneficiary meets the majority
share of QEM services to be aligned with ACOs.
The treatment group in the study was comprised of beneficiaries who were attributed to
ACOs, and the control group included beneficiaries who were attributed to non-ACOs but who
nevertheless had at least one QEM visit with an ACO provider in the year. Limiting the control

11
group to beneficiaries who have at least seen an ACO provider for primary care services
increased comparability between the treated and control groups at baseline (Supplementary
Table 2.5). In order to be considered as having pre-existing SMI conditions, beneficiaries need to
be identified with schizophrenia, bipolar disorder, or depressive disorder diagnoses using the
CMS Chronic Condition Warehouse (CCW) algorithms before the first year of the MSSP (see
Supplementary Table 2.3 for details of identification criteria). In each year, beneficiaries with
newly developed SMI conditions after the first year of MSSP were excluded to eliminate ACO’s
potential influence on SMI diagnoses or changes in coding practices among ACO providers
(Busch et al., 2016). The primary analysis included all patients with SMIs. Secondary analysis
focused on patients with SMIs and the following comorbid conditions with significantly elevated
risks among SMI persons: cardiovascular disease (CVD) (which combined heart failure, acute
myocardial infarction, ischemic heart disease, stroke, transient ischemic attack, and atrial
fibrillation), diabetes, chronic obstructive pulmonary disease (COPD), chronic kidney disease
(CKD), and Alzheimer's disease and related disorders (ADRD). Comorbid conditions were
identified in Medicare claims using diagnosis codes and CCW algorithms in each year. A full list
of diagnosis codes for comorbid conditions is included in Supplementary Table 2.4.
The outcomes of interest were annual health care utilization and spending in inpatient,
SNF, ER, and outpatient settings. I relied on the following data files to identify the outcome
variables: the MedPAR files, the outpatient claims files, and the carrier claims files. The
MedPAR files were used to identify Medicare-covered inpatient hospital and SNF stays. The
outpatient files contained claim records of services provided by institutional outpatient providers.
The carrier claim files were used to identify physician services. Both inpatient MedPAR and
outpatient files were used to identify ER visits based on non-zero emergency room charge and

12
relevant revenue center codes (0450-0459, 0981), respectively. The Master Beneficiary
Summary files provided additional information regarding the beneficiary’s demographics,
enrollment status, and chronic conditions.
Empirical Approach
Difference-in-Difference-in-Differences
The primary estimation approach used in this study is a difference-in-difference-indifferences (DDD) analysis that estimates the differential impact of ACOs on health care
utilization and spending by SMI status. This approach assesses whether health care utilization
and spending changed more in beneficiaries with pre-existing SMIs who are attributed to ACOs
before and after program entry relative to concurrent changes in beneficiaries without SMIs
served by ACOs. I focused on early cohorts entering the MSSP in 2012 and 2013 because of
more comparable beneficiary characteristics between the treated and control groups at baseline
and the fact that early cohorts allowed me to examine at least three years after program entry.1
For each of the outcome measures, I estimated linear regression models at the beneficiary-level
using the following specification for ACO’s differential impact on service use and expenditure
by beneficiary’s SMI status:
�!" = � + �#��!" + �$��! + �%��! + �#(��!" × ��!)
+ �$(��!" × ��!) + �%(��! × ��!)
+ �(��!" × ��! × ��!) + �" + �! + �&'' + ��!" + �!"
1 Later cohorts entering in 2014 and 2015 were more dissimilar in baseline beneficiary characteristics between the
treatment and control groups and showed significant pre-trends in event studies.

13
where �!" represents the annual health care utilization or spending in inpatient, SNF, ER, or
outpatient settings for beneficiary � attributed to ACO provider or non-ACO provider in year �.
��! is an indicator of treatment intensity according to ACO alignment length. ��!" is an
indicator for post-treatment years based on year of initial ACO alignment (�� has a value of 1
in years 2012-2015 for the 2012 cohort, and in years 2013-2015 for the 2013 cohort). ��! is an
indicator for having pre-existing SMI conditions. �!" is a vector of individual characteristics,
including age, gender, race/ethnicity, dual enrollment in Medicaid, disability status, end-stage
renal disease (ESRD) status, and indicators for the presence of chronic conditions defined in the
CCW (except for depression). The model includes individual fixed effects �! and year fixed
effects �" to account for time-invariant individual heterogeneity and common time trends. In
addition, hospital referral region (HRR) fixed effects �&'' are included to capture systematic
regional variations in supply and demand of health care services that are irrelevant to ACO
performance. The coefficient of the triple interaction term � captures the differential effect of
ACOs on beneficiaries with SMI conditions before and after program entry as compared to those
without SMI conditions. The standard errors for the model parameters are clustered at the
organization level, with ACO providers clustering at the ACO number level and non-ACO
providers clustering at the TIN level.
Event Study
In addition to the DDD analysis, event study is used as a supplementary approach to
directly test the parallel trends assumption and examine how the treatment effects of ACOs
evolve over time among beneficiaries with pre-existing SMIs versus those without. A series of
lead and lag years are created to reflect relative years to the initial ACO alignment, leaving the

14
year prior to the initial ACO alignment as the reference year. The coefficients on the leads are
tested for pre-trends, and the coefficients on the lags are interpreted as dynamic treatment effects
of ACOs. As in the DDD models, I focused on early cohorts entering the program in 2012 or
2013.2 For each of the outcome measures, I estimated the following event study specification
separately for beneficiaries with pre-existing SMIs and their SMI-free counterparts:
�!" = � + > �( (Treat!
%
()*+
× ��() + �" + �! + �&'' + ��!" + �!"
Where �!" is the annual health care utilization or spending by health care setting for beneficiary �
attributed to ACO provider or non-ACO provider in year �. Treat! is an indicator of ACO
alignment regardless of entry cohorts. ��( is vector of indicators of the number of years before
or after the initial ACO alignment (��, is the initial alignment year, and ��*# is one year
prior to the initial alignment and is omitted as the reference year). �!" is a vector of beneficiary
characteristics, including age, gender, race/ethnicity, Medicaid dual eligibility status, disability
status, ESRD status, and presence of CCW chronic conditions (excluding depression).
Beneficiary fixed effects �!, year fixed effects �", and HRR fixed effects �&'' are also included.
The standard errors are clustered at the provider level as before. The coefficients on the lead and
lag years, represented by Σ-)*+ *$ �-(��! × ��-) and Σ(),
% �((��! × ��(), capture the
effects of ACOs on outcome measures in relation to the reference year. The differential impact of
ACOs over time by SMI status is examined by comparison of event study coefficients between
SMI and non-SMI groups.
2 The 2012 cohorts entered in April or July 2012, and the 2013 cohorts entered in January 2013. Here I estimated
event study models for 2012-2013 cohorts regardless of treatment timing. In supplementary analysis, I examined
treatment effect heterogeneity by estimating event study models separately for 2012 and 2013 cohorts and verified
that the parallel trends and dynamic treatment effects were consistent across different entry years.

15
Falsification Tests
One concern of the empirical approaches described above is selection bias. My analysis
assumes that ACOs’ differential effects on health care utilization and spending are not
systematically related to unobserved provider and beneficiary characteristics affecting ACO
assignment status. However, non-random ACO participation and selection of beneficiaries into
ACOs are reasonable concerns. Because the formation of ACOs is voluntary, more capable and
efficient providers are more likely to participate in MSSP (Chukmaitov et al., 2019; Colla,
Lewis, Tierney, et al., 2016). Meanwhile, beneficiary alignment is affected by clinical and nonclinical patient characteristics. There is suggestive evidence that providers within ACOs may
systematically favor alignment of white and healthier beneficiaries (Epstein et al., 2014; Yasaitis
et al., 2016).
To ensure that the differential impact estimated by the DDD models was not reflecting
biased beneficiary assignment algorithms which may systematically correlate with ACO
alignment, I conducted falsification tests by treating non-ACO providers as hypothetical ACO
providers in the absence of MSSP incentives. In particular, I randomly assigned ACO versus
non-ACO status among providers who were never part of MSSP ACOs during 2012-2015 and
re-estimated DDD models. In random assignment of hypothetical ACO providers, the ratio of
ACO to non-ACO NPIs each year was kept the same as before. I performed the same sample
selection and beneficiary assignment as in the primary analysis. The purpose of the falsification
tests was to eliminate the possibility that alignment algorithms and unobserved beneficiary
characteristics predictive of ACO alignment were driving the differential effects. If alignment
algorithms and associated beneficiary characteristics are not systematically correlated with

16
treatment status and outcomes, I should find no relationship between alignment with hypothetical
ACO providers and differential changes in health care utilization and spending in the absence of
MSSP incentives.
Sensitivity Analysis
The primary analysis assumes continuous treatment since beneficiary’s initial ACO
alignment. However, not all beneficiaries attributed to ACOs were continuously aligned. I tested
robustness of results to this assumption by excluding individuals who were discontinuously
aligned from the treatment group. The differential treatment effects of ACOs were expected to be
stronger among beneficiaries who were continuously aligned to ACOs.
RESULTS
Sample Composition by ACO Alignment
The primary analysis in this study included 5,593,416 beneficiary-years from 2008-2015.
On average, 17.55% beneficiaries were attributed to MSSP ACOs annually in 2012-2015 which
was consistent with existing ACO literature (McWilliams, Gilstrap, et al., 2017; McWilliams et
al., 2020; McWilliams et al., 2018). Figure 1 shows that among beneficiaries who have ever seen
an ACO provider for QEM services, 31.48% were never aligned to ACOs, 11.16% were aligned
with ACO providers for one year, 13.74% were aligned for two years, 26.69% were aligned for
three years, and 16.93% were aligned for four years in 2012-2015. Not all beneficiaries who
have ever been attributed to ACO providers were continuously aligned (Figure 2). Over a half
(53.58%) of ACO-attributed beneficiaries were continuously aligned while 46.42% were
discontinuously aligned (i.e., not aligned with ACOs for some years after the initial alignment

17
year). The majority of (83.71%) ACO-attributed beneficiaries were either continuously aligned
or aligned again after the initial ACO alignment. In terms of years of ACO alignment,
beneficiaries were most likely to be aligned to ACOs for three consecutive years in 2013-2015
(28.87%), followed by four-year consecutive alignment in 2012-2015 (24.71%).
Beneficiary Characteristics
Table 1 shows baseline characteristics between ACO-aligned and non-aligned
beneficiaries by SMI status. Regardless of SMI status, those aligned to ACO providers tended to
slightly older, less black, less Hispanic, more Asian, and less likely to be disabled, dually
enrolled in Medicaid, and having ESRD. The ACO-aligned beneficiaries were also relatively
healthier, with fewer CCW chronic conditions and lower prevalence of diabetes, CVD, COPD,
CKD, ADRD, and hypertension. In terms of baseline health care utilization and spending, those
aligned to ACOs had fewer inpatient, SNF, ER, and outpatient visits, shorter length of stay
during inpatient and SNF admissions, and lower health care expenditures in all settings.
Consistent with prior literature, these results suggested that the MSSP may be dominated by
providers treating whiter, healthier, and less disadvantaged populations (Epstein et al., 2014;
Yasaitis et al., 2016). Comparison between SMI and non-SMI beneficiaries confirmed many
findings in existing literature: those with SMIs were younger and more likely to be female,
Hispanic, disabled, and dually enrolled in Medicaid compared to those without SMIs. The
prevalence of almost all chronic conditions was substantially higher among SMI persons. In
particular, the prevalence of ADRD more than tripled among SMI persons relative to their
counterparts. The only disease where SMI persons exhibited lower prevalence rate was cancer,

18
which is a combined measure including breast cancer, colorectal cancer, prostate cancer, lung
cancer, and endometrial cancer.
Supplementary Tables 2.6A and 2.6B consider whether among all beneficiaries who were
ever aligned to ACOs, those aligned to ACOs for longer period of time differed from those
aligned for shorter time. Among both SMI and non-SMI beneficiaries, longer ACO alignment
was associated with slightly higher prevalence of chronic conditions (except for ADRD) and
lower service use in the pre-ACO period. Additional summary statistics by SMI diagnosis are
provided in Supplementary Table 2.7 to explore heterogeneity of SMIs. Despite some noticeable
differences across SMI diagnoses—depressive disorder outweighed schizophrenia and bipolar
disorder in number, featuring beneficiaries who were older with higher rates of cancer and
CVD—those aligned to ACOs, again, were less sick and utilized health care services less
frequently regardless of specific SMI diagnosis.
Differential Effects of ACOs on Health Care Utilization by SMI Status
The differential impact of MSSP ACOs on health care utilization by settings among
beneficiaries with pre-existing SMIs versus those without SMIs are summarized in Table 2. The
DDD models estimate statistically significant effects of ACOs on the number of inpatient
admissions and length of stay, SNF stays and length of stay, and ER visits, with even larger
decrease for SMI beneficiaries in inpatient length of stay, number of SNF stays, and SNF length
of stay. In particular, ACOs reduced annual SNF stays by 0.007 more stays (p<0.01, for a total
reduction of 0.012 stays or -15.3% of the pre-ACO mean) among attributed SMI beneficiaries
compared to attributed non-SMI beneficiaries. In addition, ACOs also significantly cut average
length of hospital and SNF stays for attributed SMI beneficiaries, more so than non-SMI

19
beneficiaries. Compared to non-SMI beneficiaries assigned to ACOs who on average
experienced a reduction of 0.03 days in length of hospital stay and a reduction of 0.16 days in
length of SNF stay, those with SMIs showed further reductions in hospital and SNF days, by
0.10 days (p<0.01, for a total reduction of 0.14 days or -5.1%) and 0.35 days (p<0.01, for a total
reduction of 0.51 days or -18.9%), respectively.
Table 3 shows the DDD estimates of the impact of ACOs on health care utilization
conditional on having selected comorbid chronic conditions. In other words, I am comparing
ACO’s effects on those with SMIs and comorbid conditions versus those with the same
comorbid conditions but no SMIs. Over a half of the study population have more than one
chronic disease. The sample distribution in terms of number of selected chronic conditions is
displayed in Supplementary Table 2.8, which shows the extent of double counting across disease
groups. In summary, even larger effect sizes of ACOs were observed in samples with comorbid
chronic conditions. The DDD models estimate statistically significant effects of ACOs on the
number of inpatient admissions and days, SNF stays and days, and ER visits, with larger
decrease for SMI beneficiaries in inpatient length of stay, number of SNF stays, and SNF length
of stay across comorbid chronic diareses. Specifically, those with SMI and comorbid ADRD had
the biggest reduction, both in absolute terms and in percent change, in number of inpatient stays
(-4.2%), length of inpatient stay (-7.6%), number of SNF stays (-15.3%), and number of ER
visits (-2.9%).
The event study results of ACO’s impact on health care utilization are plotted in Figure 3.
For each health care use measure, estimates for SMI and non-SMI persons are depicted on the
same graph for comparison. All coefficients are estimated in relation to the reference year, the
year prior to the initial ACO alignment. The parallel trends assumption is assured for inpatient

20
and SNF stays and length of inpatient stay among both SMI and non-SMI beneficiaries since the
estimated coefficients on the leads of treatment are not statistically different from zero. There is
some weak indication of differential pre-trends for SNF days, ER visits, and outpatient visits.
However, the 95% confidence intervals on the leads of treatment overlapped considerably
between the non-SMI and the SMI sample and the trends are very small in magnitude compared
to the strong treatment effects in the post-ACO period. Overall, there is clear evidence of ACO
treatment effects on service utilization in all settings. Annual inpatient, SNF, ER, and outpatient
visits per beneficiary as well as length of inpatient and SNF stay fell substantially one year after
the initial alignment and continued declining as more time passed. Comparison between SMI and
non-SMI beneficiaries confirms the DDD findings that ACOs cut inpatient, SNF, and ER visits
among beneficiaries with pre-existing SMIs, more so than non-SMI beneficiaries.
Differential Effects of ACOs on Health Care Spending by SMI Status
Table 4 details DDD estimates of ACO’s differential effects on health care spending,
expressed in $2015 USD, by health care settings for beneficiaries with pre-existing SMIs
compared to those without SMIs. Among early cohorts without SMIs, ACOs reduced inpatient
spending, on average, by $107.74 (-4.2%, p<0.01), SNF spending by $84.18 (-33.5%, p<0.01),
ER spending by $11.30 (-3.9%, p<0.01), and outpatient spending by $31.58 (-1.8%, p<0.05) per
beneficiary. ACO-aligned SMI beneficiaries experienced even larger reduction in inpatient
spending by $150.24 (p<0.01), generating a total reduction of $257.98 per beneficiary or -4.5%
of the pre-ACO mean. ACOs also reduced SNF and ER spending among aligned SMI
beneficiaries, more so than non-SMI beneficiaries, by $149.83 (p<0.01, for a total reduction of

21
$234.01 or -18.2%) and $28.70 per beneficiary (p<0.01, for a total reduction of $40.00 or -
5.0%), respectively.
Table 5 demonstrates the effects of ACOs on health care spending conditional on having
selected comorbid diseases. Once again, DDD models estimate statistically significant and larger
effect sizes of ACOs on health care expenditures among samples with comorbid chronic
conditions, both in absolute terms and in percent change. This is especially evident in inpatient,
SNF, and ER expenditures. Compared to individuals without SMIs who were assigned to ACOs,
ACO-assigned SMI individuals experienced larger reductions in inpatient, SNF, and ER
expenditures across comorbid chronic diseases. Individuals with SMI and comorbid ADRD
exhibited the biggest reduction in inpatient spending (-6.6%) and ER spending (-6.9%).
Individuals with SMI and comorbid ADRD, CKD, and diabetes experienced the largest decrease
in SNF expenditure (about -19.0%).
Event study results for health care spending outcomes are shown in Figure 4. Estimates
for SMI and non-SMI samples are plotted together for each outcome measure. All coefficients
are estimated relative to one year before to the initial ACO alignment. The event study confirms
the parallel trend assumption for inpatient spending among both SMI and non-SMI beneficiaries.
There is some weak indication of differential pre-trends between the treated and control groups
in SNF and ER spending. However, the 95% confidence intervals of the SMI and non-SMI
samples mostly overlap in the pre-ACO period, and the differential pre-trends are very small in
magnitude relative to the strong treatment effects observed in post-ACO years. Outpatient
spending also shows some pre-trends in the non-SMI sample, but once again, the magnitude is
very small in relation to post-alignment treatment effects. The event studies corroborate the DDD
findings shown in Table 4. Specifically, annual inpatient, SNF, ER, and outpatient spending per

22
beneficiary fell significantly starting from one year after the initial ACO alignment, with larger
savings generated with longer alignment. Moreover, comparison between SMI and non-SMI
persons reveals bigger reductions in SNF, ER, and inpatient spending but smaller decrease in
outpatient spending among SMI beneficiaries who were aligned to ACOs relative to their
counterparts.
Falsification Tests
Results of the falsification tests are demonstrated in Supplementary Tables 2.9 and 2.10.
Notably, the DDD analyses revealed no relationship between hypothetical ACO alignment and
differential changes in most health care utilization and spending outcomes between SMI and
non-SMI beneficiaries in 2008-2015. This suggests that there is no clear evidence of bias
resulting from beneficiary assignment algorithms or corresponding beneficiary characteristics. A
few weakly significant relationships were observed, but they were opposite in direction from the
primary analysis. They suggest that, if anything, I might have underestimated the additional
reduction caused by ACOs in health care utilization and spending among SMI beneficiaries
compared to non-SMI beneficiaries in certain outcome measures. These findings provided
evidence that the ACO alignment algorithm and the beneficiary-specific factors affecting ACO
alignment were not driving the differential effects observed in the primary analysis.
Sensitivity Analyses
Restricting treatment groups to those who were continuously aligned with ACO providers
generated estimates in the same direction as in the primary analysis. However, the effect sizes
were much larger for most of the health outcomes between SMI and non-SMI beneficiaries who

23
were continuously aligned. The results suggest that ACOs may be more capable of substantially
impacting health care utilization and spending with longer alignment time (Supplementary
Tables 2.11 & 2.12).
DISCUSSION
Persons with SMIs have historically suffered from high health care costs and poor health
outcomes, due in part to providers’ lack of financial incentives to work on care coordination to
manage their many comorbid chronic conditions. Alternative payment models, such as the
MSSP, have the potential to reduce excessive health care utilization and improve health
outcomes for this vulnerable population because of the inclusion of global cost and quality
targets into provider payments. It’s important to understand how providers respond to newly
introduced incentives in alternative payment models in caring for the most vulnerable
populations to ensure that the models benefit the populations that need them most and don’t
produce unintended harm (Joynt Maddox, 2018). Taking the MSSP ACOs as an example, the
current study is the first to examine whether and the extent to which participating providers
differentially improved health outcomes for those with SMIs, especially those with both SMIs
and other comorbid chronic conditions.
My analyses suggested that ACOs generated improvements in most health outcomes
among aligned SMI beneficiaries that were at least on par with, if not greater than, those among
non-SMI beneficiaries. This is reflected in decreases in all-cause inpatient, SNF, and ER visits.
The differential decline was even more pronounced among SMI patients with comorbid chronic
conditions. In terms of health care settings, ACOs seemed to be most successful in cutting SNF
expenditure among SMI patients (-16.4% to -18.7%). This is consistent with existing literature

24
that ACO’s savings are largely driven by reductions in post-acute SNF utilization (McWilliams,
Gilstrap, et al., 2017). Additionally, ACOs also demonstrated moderate success in reducing
inpatient (-4.2% to -6.6%) and ER spending (-4.6% to -6.9%). In terms of comorbid chronic
diseases, ACOs seemed to manage comorbid ADRD better than other conditions (-6.6% in
inpatient spending, -18.7% in SNF spending, and -6.9% in ER spending).
3
One possible reason for larger effect sizes among SMI beneficiaries is the diminishing
marginal returns of health care, according to which greater improvements in health outcomes
could be achieved among SMI beneficiaries who were sicker and had more unmet health care
needs at baseline than non-SMI beneficiaries. Therefore, even provision of simple services in
primary care settings, such as risk factor screening and routine disease monitoring, could
generate substantial health gains and reduce acute and post-acute care use (Joynt Maddox, 2018).
Conceptually, one of the major mechanisms for ACO providers to achieve global cost reductions
is to shift health services from expensive settings (e.g., inpatient, SNF, and ER) to cheaper
settings (e.g., outpatient). Therefore, given the MSSP incentives, I expected to see decreased
inpatient, SNF, and ER use that was accompanied by increased outpatient use. However, such an
increase in outpatient care use was not observed. One possible reason is that I didn’t differentiate
types of outpatient services in this study, as it’s plausible for ACOs to increase uses of
preventative services while limiting unnecessary procedural and diagnostic services in outpatient
settings. Future studies should look more closely into different types of outpatient services to
better understand ACO behaviors.
3 I tested and ruled out the possibility that ACO’s larger reductions in inpatient, SNF, and ER utilization and
spending in ADRD sample were reflecting lower diagnosis rates of ADRD. The possibility of receiving a new
ADRD diagnosis between 2012-2015 didn’t statistically differ between beneficiaries aligned to ACOs versus nonACOs. The same was true for other chronic conditions.

25
The study has several limitations. First, the study focused on how the effects of ACOs on
health care utilization and spending. Quality of care, despite being an essential element in the
MSSP, was not directly examined in the current study.
Second, my empirical approaches could not isolate the mechanisms within ACOs that are
responsible for the differentially improved health outcomes that I found among SMI
beneficiaries. While it’s unlikely that system-wide changes within ACOs, such as investments in
electronic health records and administrative changes, would disproportionately benefit SMI
beneficiaries, several mechanisms are still possible in explaining the findings. It remains unclear
whether greater reductions among SMI beneficiaries in number of SNF stays and length of
inpatient and SNF stays reflect strengthened preventative care, improved screening for disease
risk factors, better care coordination across providers, better integration of medical and
behavioral care, or other health care delivery related factors that are particularly beneficial for
this vulnerable population. However, identifying the mechanism is not necessary to draw policyrelevant conclusions about the impact of MSSP ACOs on SMI populations.
While the falsification tests provided evidence that the differential effects of ACOs were
not systematically correlated with assignment algorithms or beneficiary predictors of treatment, I
cannot rule out the possibility that the differential impact was correlated with unobserved
provider-specific characteristics simultaneously affecting provider’s voluntary participation in
the MSSP and health care utilization and spending of attributed beneficiaries.
Additionally, since the identification of pre-existing SMI conditions was based on claims,
all identified beneficiaries in the study had at least one visit during the reference period while
those who didn’t access any services were unintentionally excluded. However, this concern is
alleviated by the fact that the majority (72.9%) of SMI patients aged 50 and above received some

26
type of mental health services in the past year (Substance Abuse and Mental Health Services
Administration (SAMHSA), 2020).
Finally, my estimates are based on early MSSP ACOs. Thus, the results may not directly
generalize to ACOs entering the MSSP in more recent years which may have different provider
characteristics and serve populations with different patient characteristics.
In conclusion, analyses in this study quantified the ACO-related improvements in a
variety of health care utilization and spending outcomes among SMI beneficiaries who have
historically suffered from excessively high health care costs and poor health outcomes. This
study suggested that the MSSP model may incentive providers to deliver coordinated care and
improve disease management for clinically vulnerable populations with multiple comorbid
conditions. Findings from this study could also inform design of future alternative payment
models to better align providers’ financial incentives with benefits for the most vulnerable
populations.

27
31.48%
11.16%
13.74%
26.69%
16.93%
0 year
1 year
2 years
3 years
4 years
4.73%
11.57%
5.78%
0.89% 1.52%
6.79% 5.08% 4.90%
3.00% 2.17%
28.87%
24.71%
0.00%
5.00%
10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
a2012
a2013
a2012_2013
a2012_2014
a2012_2015
a2013_2014
a2013_2015
a2012_2013_2014
a2012_2013_2015
a2012_2014_2015
a2013_2014_2015
a2012_2013_2014_2015
TABLES & FIGURES
Figure 1. Sample Composition by ACO Alignment Length, 2012-2015
Note: Sample is restricted to beneficiaries who have ever seen an ACO provider for QEM services.
Figure 2. Sample Composition by ACO Alignment Years, 2012-2015
Note: Sample is restricted to beneficiaries who were ever aligned to ACO providers.

28
Table 1. Beneficiary Characteristics by ACO Alignment and SMI Status, 2008-2011
SMI No SMI
ACO-Aligned
(1)
Not Aligned
(2)
ACO-Aligned
(3)
Not Aligned
(4)
N (Beneficiary-years) 552,896 333,944 1,363,276 546,592
Demographic Characteristics
Age 67.29 66.38 73.15 73.08
Male (%) 31.63 30.15 46.41 44.79
Race/Ethnicity (%)
White 89.50 87.32 89.09 87.50
Black 6.91 8.34 6.81 8.06
Hispanic 1.51 1.85 0.84 1.00
Asian 0.78 0.72 1.56 1.39
Eligibility Status (%)
Medicaid Dual Eligibility 30.80 34.86 9.92 10.94
Disability 39.98 44.16 12.10 14.16
End-stage Renal Disease 0.88 1.06 0.56 0.84
Chronic Conditions (%)
Cancer a 12.49 12.61 14.73 15.63
CVD b 57.99 60.25 53.54 56.29
Diabetes 36.55 39.57 30.91 33.88
CKD c 19.65 21.88 15.00 17.74
ADRD d 14.95 17.32 4.27 5.06
COPD e 29.39 33.20 17.61 20.88
Hypertension 78.29 79.47 77.85 78.72
Hyperlipidemia 76.73 76.52 77.96 78.02
CCW Conditions (#) 7.38 7.73 5.82 6.21
Annual Healthcare Utilization
Inpatient Stays 0.448 0.522 0.176 0.208
Inpatient Days 2.689 3.187 0.734 0.896
SNF Stays 0.080 0.092 0.020 0.023
SNF Days 2.706 3.338 0.464 0.557
ER Visits 1.394 1.738 0.450 0.546
Outpatient Visits 6.279 6.790 3.566 4.060
Annual Healthcare Spending
(2015$)
Inpatient Spending 5,761.10 6,580.19 2,457.86 2,896.34
SNF Spending 1,188.22 1,408.46 238.46 283.10
ER Spending 805.49 956.53 276.98 331.95
Outpatient Spending 2,854.33 3,234.16 1,696.05 1,982.39
Notes: Sample is restricted to beneficiaries who visited an ACO provider for QEM services. Beneficiaries who
develop SMI later than 2012 are excluded. Values reflect the average across pre-treatment period (2008-2011). All
spending is converted to 2015$. a Cancer includes breast cancer, colorectal cancer, prostate cancer, lung cancer, and
endometrial cancer. b CVD includes heart failure, acute myocardial infarction, ischemic heart disease, stroke,
transient ischemic attack, and atrial fibrillation. c CKD: chronic kidney disease. d ADRD: Alzheimer's disease and
related dementias. e COPD: Chronic obstructive pulmonary disease.

29
Table 2. Effect of ACOs on Health Care Utilization
Inpatient
Stays
Inpatient
Length of
Stay
SNF
Stays
SNF
Length of
Stay
ER Visits Outpatient
Visits
(1) (2) (3) (4) (5) (6)
Full Sample (N= 5,586,700)
Post × Treat -0.00582*** -0.03331*** -0.00472*** -0.15924*** -0.01512*** -0.0232
(0.00125) (0.00874) (0.00061) (0.02251) (0.00275) (0.03756)
Post × Treat × SMI -0.00343 -0.10260*** -0.00748*** -0.35363*** 0.00095 0.02629
(0.00260) (0.03364) (0.00124) (0.05218) (0.00872) (0.02447)
Percent Change -2.1% -5.1% -15.3% -18.9% -1.0% 0.05%
Pre-ACO Mean 0.45 2.69 0.08 2.71 1.40 6.28
Notes: Sample is restricted to beneficiaries who have ever visited an ACO provider for QEM services. Beneficiaries
who develop SMI later than 2012 are excluded. Treated sample includes only early cohorts. DDD estimates are
shown, adjusted for time-varying individual covariates and individual, year, and HRR fixed effects. Adding HRR
fixed effects further excludes 6,716 observations whose zip codes cannot be matched to an HRR. Robust standard
errors clustered at the provider level are in parentheses, significance at the *** 0.01, ** 0.05, and * 0.1 level.

30
Table 3. Effect of ACOs on Health Care Utilization by Comorbid Chronic Conditions
Inpatient
Stays
Inpatient
Length of
Stay
SNF
Stays
SNF
Length of
Stay
ER Visits Outpatient
Visits
(1) (2) (3) (4) (5) (6)
A. CVD Sample (N= 4,029,872)
Post × Treat -0.00870*** -0.04955*** -0.00587*** -0.20376*** -0.02268*** -0.03235
(0.00153) (0.01129) (0.00082) (0.03065) (0.00337) (0.04402)
Post × Treat × SMI -0.00277 -0.11373*** -0.00898*** -0.41693*** 0.00617 0.02110
(0.00335) (0.04208) (0.00163) (0.07052) (0.00989) (0.02734)
Percent Change -2.2% -5.3% -14.9% -17.9% -1.5% -0.2%
Pre-ACO Mean 0.53 3.12 0.10 3.47 1.52 6.69
B. COPD Sample (N=1,978,608)
Post × Treat -0.00821** -0.05161*** -0.00679*** -0.22605*** -0.02500*** -0.02100
(0.00254) (0.01972) (0.00107) (0.03871) (0.00528) (0.04588)
Post × Treat × SMI -0.00268 -0.14659*** -0.00920*** -0.47233*** 0.01284 0.00897
(0.00486) (0.05647) (0.00208) (0.09077) (0.01543) (0.03724)
Percent Change -1.7% -5.3% -14.5% -18.3% -1.3% -0.2%
Pre-ACO Mean 0.63 3.75 0.11 3.81 1.91 7.32
C. CKD Sample (N=2,170,200)
Post × Treat -0.01041*** -0.06955*** -0.00821*** -0.29505*** -0.02418*** -0.04650
(0.00242) (0.01907) (0.00125) (0.04551) (0.00495) (0.04980)
Post × Treat × SMI -0.00495 -0.17208*** -0.01119*** -0.54390*** -0.00337 0.04344
(0.00518) (0.05482) (0.00256) (0.10713) (0.01736) (0.03763)
Percent Change -2.3% -6.1% -14.9% -18.9% -1.6% -0.04%
Pre-ACO Mean 0.66 3.93 0.13 4.45 1.73 7.32
D. ADRD Sample (N=1,128,536)
Post × Treat -0.01598*** -0.10240*** -0.01627*** -0.65748*** -0.03707*** -0.01581
(0.00344) (0.02677) (0.00204) (0.09017) (0.00721) (0.03863)
Post × Treat × SMI -0.00917 -0.19685*** -0.00821** -0.45327*** -0.00698 -0.02836
(0.00588) (0.07197) (0.00332) (0.14767) (0.01561) (0.03787)
Percent Change -4.2% -7.6% -15.3% -17.5% -2.9% -0.7%
Pre-ACO Mean 0.60 3.95 0.16 6.33 1.54 6.45
E. Diabetes Sample (N=2,481,288)
Post × Treat -0.00780*** -0.03911** -0.00611*** -0.20993*** -0.02301*** -0.04539
(0.00204) (0.01638) (0.00096) (0.03378) (0.00358) (0.04442)
Post × Treat × SMI -0.00528 -0.14871*** -0.00950*** -0.46180*** 0.00360 0.05410
(0.00462) (0.05015) (0.00217) (0.09569) (0.01344) (0.03432)
Percent Change -2.3% -5.3% -14.2% -18.0% -1.1% 0.1%
Pre-ACO Mean 0.57 3.53 0.11 3.75 1.63 6.95
Notes: Sample includes beneficiaries diagnosed with selected chronic diseases and who have ever visited an ACO
provider for QEM services. Beneficiaries who develop SMI later than 2012 are excluded. Treated sample includes
only early cohorts. DDD estimates are shown, adjusted for time-varying individual covariates and individual, year,
and HRR fixed effects. Robust standard errors clustered at the provider level are in parentheses, significance at the
*** 0.01, ** 0.05, and * 0.1 level.

31
Figure 3. Effect of ACOs on Health Care Utilization

32
Table 4. Effect of ACOs on Health Care Spending
Inpatient
Payment
SNF
Payment
ER Payment Outpatient
Payment
(1) (2) (3) (4)
Full Sample (N= 5,586,700)
Post × Treat -107.743*** -84.184*** -11.301*** -31.579**
(21.378) (10.091) (3.131) (12.775)
Post × Treat × SMI -150.243*** -149.830*** -28.700*** 17.21
(41.802) (21.106) (6.641) (15.852)
Percent Change -4.5% -18.2% -5.0% -0.5%
Pre-ACO Mean 5,761.10 1,188.22 805.49 2,854.33
Notes: Sample is restricted to beneficiaries who have ever visited an ACO provider for QEM services. Beneficiaries
who develop SMI later than 2012 are excluded. Treated sample includes only early cohorts. All regressions are
adjusted for time-varying individual covariates and individual, year, and HRR fixed effects. Adding HRR fixed
effects further excludes 6,716 observations whose zip codes cannot be matched to an HRR. Spending is converted to
2015$. Robust standard errors clustered at the provider level are in parentheses, significance at the *** 0.01, **
0.05, and * 0.1 level.

33
Table 5. Effect of ACOs on Health Care Spending by Comorbid Chronic Conditions
Inpatient
Payment
SNF
Payment
ER Payment Outpatient
Payment
(1) (2) (3) (4)
A. CVD Sample (N= 4,029,872)
Post × Treat -154.068*** -107.293*** -17.984*** -47.031***
(26.770) (13.641) (4.231) (14.853)
Post × Treat × SMI -164.823*** -180.051*** -32.445*** 16.807
(56.517) (26.768) (8.574) (18.735)
Percent Change -4.2% -16.4% -4.6% -0.8%
Pre-ACO Mean 7657.89 1747.93 1089.52 3539.45
B. COPD Sample (N=1,978,608)
Post × Treat -162.002*** -121.308*** -23.821*** -15.087
(44.741) (17.600) (6.974) (17.905)
Post × Treat × SMI -194.550** -185.981*** -33.656*** -1.456
(77.990) (35.774) (12.999) (25.957)
Percent Change -4.5% -18.2% -5.0% -0.5%
Pre-ACO Mean 7951.35 1679.65 1153.01 3449.58
C. CKD Sample (N=2,170,200)
Post × Treat -187.069*** -145.576*** -27.875*** -53.812**
(42.708) (20.452) (6.526) (21.519)
Post × Treat × SMI -250.313*** -215.933*** -39.087*** 29.146
(93.135) (42.030) (13.648) (31.624)
Percent Change -5.0% -18.6% -5.8% -0.7%
Pre-ACO Mean 8741.44 1945.90 1162.38 3700.15
D. ADRD Sample (N=1,128,536)
Post × Treat -213.457*** -325.130*** -44.467*** -11.535
(56.609) (35.089) (9.901) (22.460)
Post × Treat × SMI -286.448*** -169.885*** -30.313** -2.799
(91.997) (59.840) (15.309) (32.185)
Percent Change -6.6% -18.7% -6.9% -0.5%
Pre-ACO Mean 7615.25 2645.36 1083.02 2931.78
E. Diabetes Sample (N=2,481,288)
Post × Treat -120.307*** -116.381*** -22.369*** -38.066**
(37.431) (15.835) (5.328) (17.209)
Post × Treat × SMI -243.348*** -190.303*** -40.599*** 25.197
(83.374) (36.546) (11.935) (28.353)
Percent Change -4.8% -18.6% -6.1% -0.4%
Pre-ACO Mean 7536.37 1638.50 1027.75 3376.86
Notes: Sample includes beneficiaries diagnosed with selected chronic diseases and who have ever visited an ACO
provider for QEM services. Beneficiaries who develop SMI later than 2012 are excluded. Treated sample includes
only early cohorts. DDD estimates are shown, adjusted for time-varying individual covariates and individual, year,
and HRR fixed effects. Spending is converted to 2015$. Robust standard errors clustered at the provider level are in
parentheses, significance at the *** 0.01, ** 0.05, and * 0.1 level.

34
Figure 4. Effect of ACOs on Health Care Spending

35
APPENDIX
Supplementary Table 2.1. Qualified Evaluation and Management Services
Description HCPCS Codes
Office and other outpatient services
New patient visit 99201-99205
Established patient visit 99211-99215
Nursing facility services
Initial new or established patient visit 99304-99306
Subsequent new or established patient visit 99307-99310
Discharge new or established patient services 99315-99316
Other new or established patient service 99318
Domiciliary, rest home, or custodial care services
New patient visit 99324-99328
Established patient visit 99334-99337
Oversight services 99339-99340
Home services
New patient visit 99341-99345
Established patient visit 99347-99350
Chronic care management service 99490
Transitional care management services 99495-99496
Wellness visits
Welcome visit G0402
Annual wellness visit G0438-G0439
Hospital outpatient clinic visit G0463
Note: The HCPCS codes are taken from the CMS Medicare Shared Savings Program Beneficiary Assignment
documentation.

36
Supplementary Table 2.2. Physician and Non-Physician Specialty Codes Used in
Beneficiary Assignment
Description Specialty Code
Primary care
General practice 01
Family practice 08
Internal medicine 11
Geriatric medicine 38
Nurse practitioner 50
Clinical nurse specialist 89
Physician assistant 97
Specialist with primary care designation
Cardiology 06
Osteopathic manipulative medicine 12
Neurology 13
Obstetrics/gynecology 16
Sports medicine 23
Physical medicine and rehabilitation 25
Psychiatry 26
Geriatric psychiatry 27
Pulmonary disease 29
Nephrology 39
Endocrinology 46
Multispecialty clinic or group practice 70
Addiction medicine 79
Hematology 82
Hematology/oncology 83
Preventive medicine 84
Neuropsychiatry 86
Medical oncology 90
Gynecologist/oncologist 98
Note: The specialty codes are taken from the CMS Medicare Shared Savings Program Beneficiary Assignment
documentation.

37
Supplementary Table 2.3. Algorithms for Identifying Serious Mental Illnesses
Diagnosis
Name
Reference
Period
Number/Typ
e of Claims ICD-10 Codes ICD-9 Codes
Schizophrenia 2 Years At least 1
inpatient
claim OR 2
other nondrug claims of
any service
type with DX
codes
DX F20.0, F20.1, F20.2,
F20.3, F20.5, F20.81,
F20.89, F20.9, F25.0,
F25.1, F25.8, F25.9 (any
DX on the claim)
DX 295.00, 295.01, 295.02,
295.03, 295.04, 295.05, 295.10,
295.11, 295.12, 295.13, 295.14,
295.15, 295.20, 295.21, 295.22,
295.23, 295.24, 295.25, 295.30,
295.31, 295.32, 295.33, 295.34,
295.35, 295.40, 295.41, 295.42,
295.43, 295.44, 295.45, 295.50,
295.51, 295.52, 295.53, 295.54,
295.55, 295.60, 295.61, 295.62,
295.63, 295.64, 295.65, 295.70,
295.71, 295.72, 295.73, 295.74,
295.75, 295.80, 295.81, 295.82,
295.83, 295.84, 295.85, 295.90,
295.91, 295.92, 295.93, 295.94,
295.95 (any DX on the claim)
Bipolar
Disorder
2 Years At least 1
inpatient
claim OR 2
other nondrug claims of
any service
type with DX
codes
DX F30.10, F30.11,
F30.12, F30.13, F30.2,
F30.3, F30.4, F30.8, F30.9,
F31.0, F31.10, F31.11,
F31.12, F31.13, F31.2,
F31.30, F31.31, F31.32,
F31.4, F31.5, F31.60,
F31.61, F31.62, F31.63,
F31.64, F31.70, F31.71,
F31.72, F31.73, F31.74,
F31.75, F31.76, F31.77,
F31.78, F31.81, F31.89,
F31.9, F33.8, F34.81,
F34.89, F34.9, F39 (any
DX on the claim)
DX 296.00, 296.01, 296.02,
296.03, 296.04, 296.05, 296.06,
296.10, 296.11, 296.12, 296.13,
296.14, 296.15, 296.16, 296.40,
296.41, 296.42, 296.43, 296.44,
296.45, 296.46, 296.50, 296.51,
296.52, 296.53, 296.54, 296.55,
296.56, 296.60, 296.61, 296.62,
296.63, 296.64, 296.65, 296.66,
296.7, 296.80, 296.81, 296.82,
296.89, 296.90, 296.99 (any DX
on the claim)
Depressive
Disorders
2 Years At least 1
inpatient,
SNF, HHA,
HOP, or
Carrier claim
with DX
codes
DX F32.0, F32.1, F32.2,
F32.3, F32.4, F32.5,
F32.89, F32.9, F32.A,
F33.0, F33.1, F33.2, F33.3,
F33.40, F33.41, F33.42,
F33.8, F33.9, F34.1 (any
DX on the claim)
DX 296.20, 296.21, 296.22,
296.23, 296.24, 296.25, 296.26,
296.30, 296.31, 296.32, 296.33,
296.34, 296.35, 296.36, 300.4,
311 (any DX on the claim)
Notes: Due to the lack of CCW algorithm for major depressive disorders, the algorithm for depressive disorders
were used instead which resulted in more individuals identified with the diagnosis. SNF refers to skilled nursing
facility; HHA refers to home health agency; HOP refers to hospital outpatient. When two claims are required, they
must occur at least one day apart. These algorithms are taken from the CMS Chronic Conditions Warehouse.

38
Supplementary Table 2.4. ICD Codes for Comorbid Chronic Diseases
Diagnosis Name ICD-10 Codes ICD-9 Codes
Heart Failure I09.81, I11.0, I13.0, I13.2, I50.1, I50.20,
I50.21, I50.22, I50.23, I50.30, I50.31,
I50.32, I50.33, I50.40, I50.41, I50.42,
I50.43, I50.810, I50.811, I50.812, I50.813,
I50.814, I50.82, I50.83, I50.84, I50.89,
I50.9
398.91, 402.01, 402.11, 402.91, 404.01,
404.03, 404.11, 404.13, 404.91, 404.93,
428.0, 428.1, 428.20, 428.21, 428.22,
428.23, 428.30, 428.31, 428.32, 428.33,
428.40, 428.41, 428.42, 428.43, 428.9
Ischemic Heart
Disease
I20.0, I20.1, I20.8, I20.9, I21.01, I21.02,
I21.09, I21.11, I21.19, I21.21, I21.29,
I21.3, I21.4, I21.A1, I21.A9, I22.0, I22.1,
I22.2, I22.8, I22.9, I23.0, I23.1, I23.2,
I23.3, I23.4, I23.5, I23.6, I23.7, I23.8,
I24.0, I24.1, I24.8, I24.9, I25.10, I25.110,
I25.111, I25.118, I25.119, I25.2, I25.3,
I25.41, I25.42, I25.5, I25.6, I25.700,
I25.701, I25.708, I25.709, I25.710,
I25.711, I25.718, I25.719, I25.720,
I25.721, I25.728, I25.729, I25.730,
I25.731, I25.738, I25.739, I25.750,
I25.751, I25.758, I25.759, I25.760,
I25.761, I25.768, I25.769, I25.790,
I25.791, I25.798, I25.799, I25.810,
I25.811, I25.812, I25.82, I25.83, I25.84,
I25.89, I25.9
410.00, 410.01, 410.02, 410.10, 410.11,
410.12, 410.20, 410.21, 410.22, 410.30,
410.31, 410.32, 410.40, 410.41, 410.42,
410.50, 410.51, 410.52, 410.60, 410.61,
410.62, 410.70, 410.71, 410.72, 410.80,
410.81, 410.82, 410.90, 410.91, 410.92,
411.0, 411.1, 411.81, 411.89, 412, 413.0,
413.1, 413.9, 414.00, 414.01, 414.02,
414.03, 414.04, 414.05, 414.06, 414.07,
414.12, 414.2, 414.3, 414.4, 414.8, 414.9
Acute
Myocardial
Infarction
I21.01, I21.02, I21.09, I21.11, I21.19,
I21.21, I21.29, I21.3, I21.4, I21.9, I21.A1,
I21.A9, I22.0, I22.1, I22.2, I22.8, I22.9
410.01, 410.11, 410.21, 410.31, 410.41,
410.51, 410.61, 410.71, 410.81, 410.91
Atrial Fibrillation I48.0, I48.1, I48.11, I48.19, I48.2, I48.20,
I48.21, I48.91
427.31
Stroke/Transient
Ischemic Attack
G45.0, G45.1, G45.2, G45.8, G45.9, G46.0,
G46.1, G46.2, G46.3, G46.4, G46.5, G46.6,
G46.7, G46.8, G97.31, G97.32, I60.xx,
I61.x, I63.xx, I66.01, I66.02, I66.03,
I66.09, I66.11, I66.12, I66.13, I66.19,
I66.21, I66.22, I66.23, I66.29, I66.3, I66.8,
I66.9, I67.841, I67.848, I67.89, I97.810,
I97.811, I97.820, I97.821
430, 431, 433.01, 433.11, 433.21, 433.31,
433.81, 433.91, 434.00, 434.01, 434.10,
434.11, 434.90, 434.91, 435.0, 435.1,
435.3, 435.8, 435.9, 436, 997.02
Chronic
Obstructive
Pulmonary
Disease and
Bronchiectasis
(COPD)
J40, J41.0, J41.1, J41.8, J42, J43.0, J43.1,
J43.2, J43.8, J43.9, J44.0, J44.1, J44.9,
J47.0, J47.1, J47.9
490, 491.0, 491.1, 491.20, 491.21, 491.22,
491.8, 491.9, 492.0, 492.8, 494.0, 494.1,
496

39
Diagnosis Name ICD-10 Codes ICD-9 Codes
Alzheimer's
Disease and
Related Disorders
or Senile
Dementia
(ADRD)
F01.50, F01.51, F02.80, F02.81, F03.90,
F03.91, F04, F05, F06.1, F06.8, G13.8,
G30.0, G30.1, G30.8, G30.9, G31.01,
G31.09, G31.1, G31.2, G94, R41.81, R54
331.0, 331.11, 331.19, 331.2, 331.7, 290.0,
290.10, 290.11, 290.12, 290.13, 290.20,
290.21, 290.3, 290.40, 290.41, 290.42,
290.43, 294.0, 294.10, 294.11, 294.20,
294.21, 294.8, 797
Chronic Kidney
Disease (CKD)
A18.11, A52.75, B52.0, C64.1, C64.2,
C64.9, C68.9, D30.00, D30.01, D30.02,
D41.00, D41.01, D41.02, D41.10, D41.11,
D41.12, D41.20, D41.21, D41.22, D59.3,
E08.21, E08.22, E08.29, E08.65, E09.21,
E09.22, E09.29, E10.21, E10.22, E10.29,
E10.65, E11.21, E11.22, E11.29, E11.65,
E13.21, E13.22, E13.29, E74.8, I12.0,
I12.9, I13.0, I13.10, I13.11, I13.2, I70.1,
I72.2, K76.7, M10.3x, M32.14, M32.15,
M35.04, N0x, N13.1, N13.2, N13.30,
N13.39, N14.0, N14.1, N14.2, N14.3,
N14.4, N15.0, N15.8, N15.9, N16, N17.0,
N17.1, N17.2, N17.8, N17.9, N18.1, N18.2,
N18.3, N18.30, N18.31, N18.32, N18.4,
N18.5, N18.6, N18.9, N19, N25.0, N25.1,
N25.81, N25.89, N25.9, N26.1, N26.9,
Q61.02, Q61.11, Q61.19, Q61.2, Q61.3,
Q61.4, Q61.5, Q61.8, Q62.0, Q62.2,
Q62.10, Q62.11, Q62.12, Q62.31, Q62.32,
Q62.39, R94.4
016.00, 016.01, 016.02, 016.03, 016.04,
016.05, 016.06, 095.4, 189.0, 189.9, 223.0,
236.91, 249.40, 249.41, 250.40, 250.41,
250.42, 250.43, 271.4, 274.10, 283.11,
403.01, 403.11, 403.91, 404.02, 404.03,
404.12, 404.13, 404.92, 404.93, 440.1,
442.1, 572.4, 580.0, 580.4, 580.81, 580.89,
580.9, 581.0, 581.1, 581.2, 581.3, 581.81,
581.89, 581.9, 582.0, 582.1, 582.2, 582.4,
582.81, 582.89, 582.9, 583.0, 583.1, 583.2,
583.4, 583.6, 583.7, 583.81, 583.89, 583.9,
584.5, 584.6, 584.7, 584.8, 584.9, 585.1,
585.2, 585.3, 585.4, 585.5, 585.6, 585.9,
586, 587, 588.0, 588.1, 588.81, 588.89,
588.9, 591, 753.12, 753.13, 753.14, 753.15,
753.16, 753.17, 753.19, 753.20, 753.21,
753.22, 753.23, 753.29, 794.4
Diabetes E08.xx, E09.xx, E10.xx, E11.xx, E13.xx 249.xx, 250.00, 250.01, 250.02, 250.03,
250.10, 250.11, 250.12, 250.13, 250.20,
250.21, 250.22, 250.23, 250.30, 250.31,
250.32, 250.33, 250.40, 250.41, 250.42,
250.43, 250.50, 250.51, 250.52, 250.53,
250.60, 250.61, 250.62, 250.63, 250.70,
250.71, 250.72, 250.73, 250.80, 250.81,
250.82, 250.83, 250.90, 250.91, 250.92,
250.93, 357.2, 362.01, 362.02, 362.03,
362.04, 362.05, 362.06, 366.41
Notes: In this paper, CVD includes any diagnosis related to heart failure, acute myocardial infarction, ischemic heart
disease, stroke, transient ischemic attack, and atrial fibrillation. The ICD codes are taken from the CMS Chronic
Conditions Warehouse.

40
Supplementary Table 2.5. Beneficiary Characteristics among Not Aligned, by Ever Seen
ACO Provider and Pre-existing SMI Status, 2008-2011
SMI No SMI
Visit ACO,
Not Aligned
(1)
Never Visit
ACO,
Not Aligned
(2)
Visit ACO,
Not Aligned
(3)
Never Visit
ACO,
Not Aligned
(4)
N 333,944 1,722,048 546,592 5,264,780
Demographic Characteristics
Age 66.38 65.21 73.08 71.58
Male (%) 30.15% 35.01% 44.79% 49.53%
Race/Ethnicity (%)
White 87.40% 86.05% 87.56% 84.97%
Black 8.34% 8.85% 8.06% 9.16%
Hispanic 1.85% 2.11% 1.00% 1.58%
Asian 0.72% 0.95% 1.39% 1.74%
Eligibility Status (%)
Medicaid Dual Eligibility 34.86% 36.47% 10.94% 13.19%
Disability 44.16% 46.78% 14.16% 17.76%
End-stage Renal Disease 1.06% 0.67% 0.84% 0.51%
Chronic Conditions (%)
Cancer a 12.61% 10.18% 15.63% 11.32%
CVD b 60.25% 50.01% 56.29% 40.68%
Diabetes 39.57% 32.49% 33.88% 25.36%
CKD 21.88% 16.06% 17.74% 10.79%
ADRD 17.32% 14.89% 5.06% 3.78%
COPD 33.20% 27.58% 20.88% 14.54%
CCW Conditions (#) 7.73 6.63 6.21 4.70
Annual Healthcare Utilization
Inpatient Stays 0.522 0.387 0.208 0.130
Inpatient Days 3.187 2.519 0.896 0.553
SNF Stays 0.092 0.068 0.023 0.014
SNF Days 3.338 2.715 0.557 0.368
ER Visits 1.738 1.328 0.546 0.390
Outpatient Visits 6.790 5.733 4.060 2.907
Annual Healthcare Spending
(2015$)
Inpatient Spending $6,580.19 $4,743.90 $4,743.90 $1,747.08
SNF Spending $1,408.46 $1,032.36 $1,032.36 $177.19
ER Spending $956.53 $670.24 $670.24 $205.42
Outpatient Spending $3,234.16 $2,422.46 $1,982.39 $1,304.44
Notes: Sample includes only beneficiaries who are never attributed to ACO providers. Those who develop SMI later
than 2012 are excluded from analysis. Values reflect the average across pre-ACO period (2008-2011). All spending
is converted to 2015$. a Cancer includes breast cancer, colorectal cancer, prostate cancer, lung cancer, and
endometrial cancer. b CVD includes heart failure, acute myocardial infarction, ischemic heart disease, stroke,
transient ischemic attack, and atrial fibrillation.

41
Supplementary Table 2.6A. Beneficiary Characteristics by ACO Alignment Length, No
Pre-existing SMI, 2008-2011
No SMI (N= 1,909,868)
Never
Aligned
(1)
Aligned 1
Year
(2)
Aligned 2
Years
(3)
Aligned 3
Years
(4)
Aligned 4
Years
(5)
N 546,592 208,188 265,260 539,044 350,784
Demographic Characteristics
Age 73.08 72.46 72.94 73.24 73.58
Male (%) 44.79% 46.03% 46.55% 46.85% 45.87%
Race/Ethnicity (%)
White 87.50% 87.28% 88.71% 89.88% 89.25%
Black 8.06% 7.63% 7.08% 6.62% 6.41%
Hispanic 1.00% 1.40% 0.99% 0.65% 0.67%
Asian 1.39% 1.74% 1.45% 1.28% 1.98%
Chronic Conditions (%)
Cancer a 15.63% 13.93% 14.64% 14.79% 15.17%
CVD b 56.29% 50.46% 53.86% 53.22% 55.61%
Diabetes 33.88% 28.97% 29.93% 30.80% 32.97%
CKD 17.74% 14.84% 15.13% 14.77% 15.35%
ADRD 5.06% 4.79% 4.75% 4.00% 4.00%
COPD 20.88% 17.48% 17.93% 17.21% 18.04%
Annual Healthcare Utilization
Inpatient Stays 0.208 0.175 0.179 0.173 0.178
Inpatient Spending $2,896.34 $2,419.03 $2,496.31 $2,429.41 $2,495.55
SNF Stays 0.023 0.021 0.021 0.020 0.019
SNF Spending $283.10 $267.48 $262.55 $233.01 $211.41
ER Visits 0.546 0.478 0.479 0.444 0.421
ER Spending $331.95 $279.43 $288.82 $273.63 $271.72
Outpatient Visits 4.060 3.645 3.549 3.620 3.449
Outpatient Spending $1,982.39 $1,783.42 $1,713.08 $1,703.15 $1,620.42
Notes: Sample is restricted to beneficiaries who have ever seen an ACO provider and have no pre-existing SMI. Of
those attributed to ACO providers, only early cohorts (i.e., initially aligned in 2012 or 2013) are included. Values
reflect the average across pre-ACO period (2008-2011). All spending is converted to 2015$.

42
Supplementary Table 2.6B. Beneficiary Characteristics by ACO Alignment Length, Preexisting SMI, 2008-2011
SMI (N=886,840)
Never
Aligned
(1)
Aligned 1
Year
(2)
Aligned 2
Years
(3)
Aligned 3
Years
(4)
Aligned 4
Years
(5)
N 333,944 104,012 118,952 207,280 122,652
Demographic Characteristics
Age 66.38 65.84 66.73 67.60 68.53
Male (%) 30.15% 31.85% 31.75% 31.94% 30.82%
Race/Ethnicity (%)
White 87.32% 87.54% 88.60% 90.20% 90.40%
Black 8.34% 8.28% 7.24% 6.52% 6.10%
Hispanic 1.85% 1.88% 1.74% 1.28% 1.34%
Asian 0.72% 0.75% 0.86% 0.72% 0.84%
Chronic Conditions (%)
Cancer a 12.61% 11.66% 12.21% 12.67% 13.17%
CVD b 60.25% 56.78% 58.77% 57.34% 59.35%
Diabetes 39.57% 36.27% 36.73% 35.98% 37.59%
CKD 21.88% 19.74% 19.74% 19.32% 20.03%
ADRD 17.32% 17.16% 16.05% 14.07% 13.51%
COPD 33.20% 30.40% 30.31% 28.51% 29.12%
Annual Healthcare Utilization
Inpatient Stays 0.522 0.492 0.477 0.424 0.425
Inpatient Spending $6,580.19 $6,193.76 $6,031.32 $5,518.30 $5,542.46
SNF Stays 0.092 0.090 0.085 0.075 0.073
SNF Spending $1,408.46 $1,397.90 $1,303.50 $1,091.94 $1,061.29
ER Visits 1.738 1.636 1.518 1.320 1.191
ER Spending $956.53 $911.53 $869.61 $759.07 $731.85
Outpatient Visits 6.790 6.515 6.264 6.250 6.141
Outpatient Spending $3,234.16 $2,967.54 $2,884.83 $2,812.64 $2,799.21
Notes: Sample is restricted to beneficiaries who have ever seen an ACO provider and have pre-existing SMI. Of
those attributed to ACO providers, only early cohorts (i.e., initially aligned in 2012 or 2013) are included. Values
reflect the average across pre-ACO period (2008-2011). All spending is converted to 2015$.

43
Supplementary Table 2.7. Beneficiary Characteristics by SMI Diagnosis and ACO
Treatment Status, 2008-2011
Notes: Sample excludes beneficiaries who have never visited an ACO provider, who develop SMI later than 2012,
or who are aligned to ACO but with first alignment year later than 2013. Values in this table reflect the average
across pre-ACO period (2008-2011). All spending is converted to 2015$. a Cancer includes breast cancer, colorectal
cancer, prostate cancer, lung cancer, and endometrial cancer. b CVD includes heart failure, acute myocardial
infarction, ischemic heart disease, stroke, transient ischemic attack, and atrial fibrillation. Comorbidity of SMI
diagnosis is high. Among beneficiaries with Schizophrenia, 32.27% and 40.87% of those aligned have comorbid
Bipolar and Depression, respectively, and 35.50% and 45.44% of those not aligned have comorbid Bipolar and
Depression, respectively. Among beneficiaries with Bipolar, 19,95% and 69.83% of those aligned have comorbid
Schizophrenia and Depression, respectively, and 21.92% and 72.61% of those not aligned have comorbid
Schizophrenia and Depression, respectively. Among beneficiaries with Depression, 5.05% and 13.95% of those
aligned have comorbid Schizophrenia and Bipolar, respectively, and 6.51 % and 16.86% of those not aligned have
comorbid Schizophrenia and Bipolar, respectively.
Schizophrenia Bipolar Depression
Aligned
(1)
Not
Aligned
(2)
Aligned
(3)
Not
Aligned
(4)
Aligned
(5)
Not
Aligned
(6)
N 34,984 25,543 56,603 41,381 283,294 178,189
Demographic Characteristics
Age 52.74 53.14 54.97 54.60 68.44 67.45
Male (%) 52.49% 49.34% 35.99% 33.46% 29.22% 27.62%
Race/Ethnicity (%)
White 76.05% 74.24% 88.12% 86.74% 89.86% 87.97%
Black 18.06% 19.63% 8.19% 9.21% 6.43% 7.70%
Hispanic 2.63% 2.82% 1.55% 1.93% 1.58% 1.94%
Asian 1.22% 1.21% 0.60% 0.42% 0.72% 0.61%
Chronic Conditions (%)
Cancer a 5.34% 6.23% 7.25% 7.58% 12.40% 12.61%
CVD b 41.58% 46.70% 45.65% 47.99% 58.66% 61.26%
Diabetes 40.91% 44.79% 35.64% 38.34% 37.15% 40.16%
CKD 16.35% 19.65% 18.52% 20.65% 20.79% 23.04%
ADRD 19.45% 22.72% 17.44% 19.50% 17.22% 19.81%
COPD 32.73% 37.69% 32.33% 37.30% 31.26% 35.49%
Annual Healthcare Use
Inpatient Stays 0.751 0.879 0.790 0.891 0.631 0.719
Inpatient Spending $8,381.9
1
$9,888.4
8
$8,652.0
8
$9,726.9
1
$8,078.9
0
$9,021.0
SNF Stays 0.099 0.117 0.093 0.114 0.122 0.136 8
SNF Spending $1,657.0
5
$1,986.1
2
$1,402.9
3
$1,795.9
5
$1,835.8
3
$2,094.2
ER Visits 2.332 2.813 2.947 3.484 1.845 2.276 5
ER Spending $1,241.5
2
$1,499.5
5
$1,432.5
8
$1,666.0
2
$1,104.4
3
$1,288.6
Outpatient Visits 7.986 8.341 8.534 8.767 7.319 7.872 8
Outpatient Spending $2,727.9
8
$3,231.6
5
$3,441.5
0
$3,724.1
9
$3,366.5
8
$3,779.1
1

44
Supplementary Table 2.8. Sample Distribution in Number of Selected Chronic Conditions
# of chronic conditions N Precent
0 1,225,071 21.93
1 1,576,828 28.22
2 1,386,178 24.81
3 904,455 16.19
4 402,915 7.21
5 91,253 1.63
Total 5,586,700 100
Notes: Sample is restricted to beneficiaries who have ever seen an ACO provider and have no pre-existing SMI. Of
those attributed to ACO providers, only early cohorts (i.e., initially aligned in 2012 or 2013) are included. Chronic
conditions include cancer, CVD, diabetes, CKD, COPD, and ADRD.

45
Supplementary Table 2.9. Falsification Test Treating Non-ACO Providers as Hypothetical
ACO Providers, Health Care Utilization
Inpatient
Stays
Inpatient
Length of
Stay
SNF
Stays
SNF
Length of
Stay
ER Visits Outpatient
Visits
(1) (2) (3) (4) (5) (6)
Full Sample (N= 4,154,489)
Post × Treat -0.00431 -0.02408 -0.00181 -0.07536* -0.00116 -0.07916*
(0.00186) (0.01258) (0.00070) (0.02605) (0.00430) (0.01911)
Post × Treat × SMI 0.00687 0.02529 0.00272 0.15852 0.02867* 0.01533
(0.00498) (0.05012) (0.00216) (0.09933) -0.00116 (0.03935)
B. CVD Sample (N= 3,036,600)
Post × Treat -0.00497 -0.02467 -0.00226 -0.09008* -0.00964 -0.12000*
(0.00247) (0.01698 (0.00094) (0.03493) (0.00550) (0.02332)
Post × Treat × SMI 0.00763 0.02089 0.00330 0.19870 0.03559* 0.06140
(0.00638) (0.06296) (0.00284) (0.13037) (0.01996) (0.04866)
C. COPD Sample (N=1,547,327)
Post × Treat -0.00396 -0.03614 -0.00301 -0.12415 0.00039 -0.13809*
(0.00429) (0.03050) (0.00167) (0.05950) (0.00977) (0.03679)
Post × Treat × SMI 0.00991 0.04990 0.00558 0.32610* 0.06829** 0.08478
(0.00924) (0.08336) (0.00405) (0.17171) (0.03025) (0.06917)
D. CKD Sample (N=1,630,085)
Post × Treat 0.00175 0.00798 -0.00246 -0.11134 0.00837 -0.08160*
(0.00414) (0.03037) (0.00160) (0.05964) (0.00890) (0.03478)
Post × Treat × SMI 0.00182 -0.05425 0.00214 0.16176 0.00964 0.05751
(0.01005) (0.09321) (0.00449) (0.20482) (0.02831) (0.07156)
E. ADRD Sample (N=875,356)
Post × Treat 0.00491 0.01278 -0.00244 -0.27139* 0.00660 -0.08307
(0.00694) (0.05143) (0.00343) (0.14727) (0.01528) (0.05205)
Post × Treat × SMI 0.00474 -0.03777 0.00315 0.39942 0.02120 0.04922
(0.01184) (0.12086) (0.00629) (0.30400) (0.03254) (0.08519)
F. Diabetes Sample (N=1,892,672)
Post × Treat -0.00405 -0.01876 -0.00179 -0.07382 -0.01247* -0.11509*
(0.00322) (0.02316) (0.00123) (0.04560) (0.00700) (0.02851)
Post × Treat × SMI 0.01206 0.02352 0.00492 0.31085* 0.05418** 0.04937
(0.00804) (0.08125) (0.00364) (0.16154) (0.02565) (0.06062)
Notes: Results in the table represent DDD estimates from falsification test that randomly assigned ACO providers
each year among providers were never associated with ACOs in 2012-2015. The falsification test performed the
same beneficiary selection, beneficiary attribution, and SMI identification procedures as in the primary analysis.
Robust standard errors clustered at the provider level are in parentheses, significance at the *** 0.01, ** 0.05, and *
0.1 level.

46
Supplementary Table 2.10. Falsification Test Treating Non-ACO Providers as Hypothetical
ACO Providers, Health Care Spending
Inpatient
Payment
SNF
Payment
ER Payment Outpatient
Payment
(1) (2) (3) (4)
A. Full Sample (N= 4,154,489)
Post × Treat -103.602 -24.932 -12.883 -29.139
(33.754) (11.371) (14.858) (32.671)
Post × Treat × SMI 77.465 77.408* 22.826 80.746*
(83.869) (37.114) (18.458) (44.622)
B. CVD Sample (N= 3,036,600)
Post × Treat -106.385 -32.681 -26.141 -62.070
(45.129) (15.198) (20.165) (43.415)
Post × Treat × SMI 75.103 98.099** 30.853 113.213**
(108.642) (49.034) (24.580) (56.106)
C. COPD Sample (N=1,547,327)
Post × Treat -128.431 -43.148 -3.365 -36.748
(77.511) (26.612) (10.518) (34.198)
Post × Treat × SMI 145.505 137.974** 34.714 107.273*
(157.992) (67.675) (23.137) (59.215)
D. CKD Sample (N=1,630,085)
Post × Treat -44.868 -35.823 3.750 28.413
(79.923) (25.935) (10.136) (36.352)
Post × Treat × SMI 11.942 107.292 -4.033 58.184
(176.531) (77.228) (24.621) (65.041)
E. ADRD Sample (N=875,356)
Post × Treat 74.040 -41.896 23.169 70.069*
(113.649) (58.595) (18.022) (42.226)
Post × Treat × SMI -136.005 133.662 -6.606 0.682
(197.350) (111.654) (30.250) (65.687)
F. Diabetes Sample (N=1,892,672)
Post × Treat -88.725 -33.395* -37.578 -75.443
(60.837) (20.238) (33.257) (69.082)
Post × Treat × SMI 142.700 140.110** 55.432 107.073
(140.486) (63.570) (37.662) (83.947)
Notes: Results in the table represent DDD estimates from falsification test that randomly assigned ACO providers
each year among providers were never associated with ACOs in 2012-2015. The falsification test performed the
same beneficiary selection, beneficiary attribution, and SMI identification procedures as in the primary analysis.
Regressions are adjusted for time-varying individual covariates and individual and year fixed effects. Robust
standard errors clustered at the provider TIN level are in parentheses. All spending is expressed in 2015$.
Significance at the *** 0.01, ** 0.05, and * 0.1 level.

47
Supplementary Table 2.11. Sensitivity Analysis of Continuously Treated, Health Care
Utilization
Inpatient
Stays
Inpatient
Length of
Stay
SNF
Stays
SNF
Length of
Stay
ER Visits Outpatient
Visits
(1) (2) (3) (4) (5) (6)
Full Sample (N=3,814,504)
Post × Treat -0.01028*** -0.06230*** -0.00621*** -0.19748*** -0.02525*** -0.04180
(0.00159) (0.01029) (0.00078) (0.03090) (0.00435) (0.05306)
Post × Treat × SMI -0.00865** -0.13787*** -0.01259*** -0.46738*** -0.00501 0.07831*
(0.00336) (0.04917) (0.00178) (0.07683) (0.01279) (0.04043)
B. CVD Sample (N=2,772,136)
Post × Treat -0.01368*** -0.08620*** -0.00756*** -0.25903*** -0.03469*** -0.04678
(0.00202) (0.01332) (0.00109) (0.04080) (0.00546) (0.05993)
Post × Treat × SMI -0.00832** -0.15050** -0.01443*** -0.54376*** -0.00434 0.06757
(0.00424) (0.05949) (0.00221) (0.09606) (0.01453) (0.04479)
C. COPD Sample (N=1,369,704)
Post × Treat -0.01529*** -0.11152*** -0.00897*** -0.31500*** -0.03277*** -0.01061
(0.00330) (0.02338) (0.00135) (0.04921) (0.00813) (0.06238)
Post × Treat × SMI -0.01048 -0.20235** -0.01717*** -0.60056*** 0.00125 0.06722
(0.00683) (0.09195) (0.00303) (0.13512) (0.02342) (0.05323)
D. CKD Sample (N=1,509,072)
Post × Treat -0.01971*** -0.13703*** -0.01160*** -0.41142*** -0.04304*** -0.07356
(0.00322) (0.02322) (0.00172) (0.06083) (0.00761) (0.06958)
Post × Treat × SMI -0.00864 -0.21293*** -0.01678*** -0.55507*** 0.00187 0.10408**
(0.00668) (0.06934) (0.00333) (0.13818) (0.02468) (0.05220)
E. ADRD Sample (N=758,504)
Post × Treat -0.02357*** -0.14388*** -0.02035*** -0.84079*** -0.06178*** -0.01656
(0.00489) (0.03428) (0.00283) (0.12153) (0.01148) (0.05557)
Post × Treat × SMI -0.01641** -0.29699*** -0.01514*** -0.55606** -0.00369 0.02754
(0.00804) (0.11228) (0.00468) (0.22279) (0.01956) (0.06268)
F. Diabetes Sample (N=1,725,296)
Post × Treat -0.01309*** -0.07519*** -0.00874*** -0.29796*** -0.03008*** -0.04949
(0.00278) (0.01969) (0.00133) (0.04571) (0.00592) (0.05909)
Post × Treat × SMI -0.01247** -0.18973** -0.01496*** -0.60091*** -0.01261 0.11518**
(0.00571) (0.08118) (0.00286) (0.12652) (0.02010) (0.05419)
Notes: Results in the table represent DDD estimates limiting to those continuously aligned to ACOs since their
initial alignment in 2012-2015. Regressions are adjusted for time-varying individual covariates and individual and
year fixed effects. Robust standard errors clustered at the provider TIN level are in parentheses. Significance at the
*** 0.01, ** 0.05, and * 0.1 level.

48
Supplementary Table 2.12. Sensitivity Analysis of Continuously Treated, Health Care
Spending
Inpatient
Payment
SNF
Payment
ER Payment Outpatient
Payment
(1) (2) (3) (4)
A. Full Sample (N=3,814,504)
Post × Treat -200.877*** -112.055*** -26.918*** -63.850***
(26.766) (12.224) (4.998) (19.198)
Post × Treat × SMI -224.323*** -219.557*** -43.015*** 46.220*
(56.498) (29.064) (8.767) (24.676)
B. CVD Sample (N=2,772,136)
Post × Treat -264.834*** -137.746*** -36.596*** -72.518***
(34.579) (16.950) (6.593) (20.483)
Post × Treat × SMI -240.600*** -253.448*** -47.414*** 34.712
(74.678) (35.830) (11.573) (28.085)
C. COPD Sample (N=1,369,704)
Post × Treat -325.712*** -167.696*** -48.035*** -33.636
(58.628) (21.587) (10.899) (27.962)
Post × Treat × SMI -288.297*** -258.979*** -49.754*** 24.453
(109.065) (48.499) (17.776) (38.271)
D. CKD Sample (N=1,509,072)
Post × Treat -373.172*** -201.075*** -57.980*** -86.057***
(56.453) (27.324) (10.600) (32.380)
Post × Treat × SMI -298.981** -290.178*** -44.758** 50.731
(129.787) (54.403) (19.128) (45.883)
E. ADRD Sample (N=758,504)
Post × Treat -304.947*** -407.554*** -76.993*** -20.797
(77.874) (45.890) (15.823) (30.751)
Post × Treat × SMI -417.512*** -253.535*** -44.267** 30.031
(134.630) (80.223) (21.500) (44.469)
F. Diabetes Sample (N=1,725,296)
Post × Treat -224.838*** -156.118*** -36.596*** -64.412**
(51.153) (20.378) (8.214) (25.799)
Post × Treat × SMI -354.535*** -283.542*** -68.318*** 41.580
(105.210) (49.392) (16.299) (41.925)
Notes: Results in the table represent DDD estimates limiting to those continuously aligned to ACOs since their
initial alignment in 2012-2015. Regressions are adjusted for time-varying individual covariates and individual and
year fixed effects. Robust standard errors clustered at the provider TIN level are in parentheses. All spending is
expressed in 2015$. Significance at the *** 0.01, ** 0.05, and * 0.1 level.

49
Chapter 3: Impact of Medicare Shared Savings Program on Health Care Spending and
Utilization Using Regression Discontinuity Design
INTRODUCTION
In 2012, the Center for Medicare and Medicaid Services (CMS) implemented the
Medicare Shared Savings Program (MSSP) as part of a broader reform to move from volumebased to value-based provider payment. The MSSP creates global incentives for voluntarily
formed accountable care organizations (ACOs)—groups of physicians, hospitals, and other
health care providers—to reduce health care spending and improve quality of care for assigned
Medicare fee-for-service (FFS) beneficiaries. Starting with 220 ACOs that joined the program in
2012 and 2013, the MSSP expanded over time with 483 ACOs serving 11 million Medicare FFS
beneficiaries in 2022 (CMS, 2022). Under the MSSP contract, participating ACOs are eligible
for financial rewards from CMS if total Medicare spending of assigned beneficiaries falls
considerably below a financial benchmark while clearing a set of quality metrics. The financial
benchmark for an ACO is determined by its spending levels during a baseline period prior to
joining the MSSP and updated each year according to national Medicare spending growth (CMS,
2014). The MSSP offers multiple tracks of participation and during the first three years of
program implementation, almost all ACOs (99%) chose to participate in the track 1 MSSP,
which are one-sided contracts with no financial penalty if total expenditures exceed benchmarks
(CMS, 2015).
Existing literature on the impact of early cohorts of MSSP ACOs all adopted a
difference-in-differences framework and found that ACOs achieved very modest reductions in
total Medicare spending with no clear evidence of deteriorated quality of care for assigned

50
beneficiaries (McWilliams, 2016; McWilliams et al., 2016; McWilliams et al., 2020;
McWilliams et al., 2018). Comparing beneficiaries served by ACOs to those served by non-ACO
providers, prior research found that by 2015, the MSSP on average resulted in a 3.1% reduction
in per-beneficiary Medicare spending among ACOs that entered the program in 2012. The
reduction in total Medicare expenditures was smaller, about 1.4%, among ACOs that entered in
2013 (McWilliams et al., 2018). Further analysis showed that reductions in expenditures varied
by health care settings and ACO’s organizational structure. Specifically, reductions were
concentrated among physician-group ACOs and driven by lower use of care in acute, post-acute,
and hospital outpatient settings (McWilliams, Gilstrap, et al., 2017; McWilliams et al., 2016;
McWilliams et al., 2018).
Ever since the introduction of the ACO model, critics have pointed to the possibility that
the formation of ACOs may not random (Anderson et al., 2014; Lewis et al., 2012; Pollack &
Armstrong, 2011). Because participation in the ACOs is voluntary, providers with greater
capacity and serve healthier patients are more likely to participate in the MSSP. Indeed, several
previous studies reported that the formation of early ACOs is associated with greater managed
care penetration, prior participation in other alternative payment models, urban locations, and
lower poverty rates in the region (Chukmaitov et al., 2019; Colla, Lewis, Tierney, et al., 2016;
Lewis et al., 2013). Consistent with these results, another study found that physicians were
significantly less likely to join in ACOs in areas with more vulnerable populations, including
racial minorities (Yasaitis et al., 2016). There is also suggestive evidence that compared to
patients not served by ACOs, those assigned to Medicare ACOs had higher incomes and were
less likely to be black, disabled, or dually covered by Medicaid (Epstein et al., 2014). These
differences reflect the non-random formation of ACOs. Therefore, identifying the causal impact

51
of ACOs requires an ability to net out differences on observable and unobservable characteristics
that result from selection which previous empirical methods might fail to do so. Additionally,
sample selections in prior studies generally included a substantial share of beneficiaries for
whom there may be very little room for ACOs to reduce spending and improve care
coordination. For example, this includes beneficiaries who relied on only one provider for their
primary care services.
To address these issues, in this study we implement a regression discontinuity approach
to identify the causal effects of early cohorts of MSSP ACOs on health care spending and
utilization among beneficiaries seeing more than one provider for primary care services. Our
approach relies on the observation that some beneficiaries can be just marginally aligned to
ACOs, while others just marginally unaligned, based on the beneficiary’s distribution of primary
care service expenditures across provider organizations. With this approach, the only difference
governing the aligned and unaligned beneficiaries are the incentives introduced by the ACOs.
METHODS
Data and Study Population
This study used Medicare administrative claims and enrollment data for annual random
20% samples of FFS beneficiaries from 2008 to 2015. For each year, the study population
included Medicare beneficiaries who were continuously enrolled in FFS Parts A and B during
the study period, not enrolled in Medicare Advantage, living in the U.S., and actively seeking
primary care services from ACO providers participating in the MSSP. In each year, we excluded
beneficiaries who did not have at least one qualified evaluation and management (QEM) service
with a physician at an ACO because it’s theoretically impossible for them to be attributed to

52
ACOs. Among this eligible sample, we performed retrospective beneficiary assignment
following the CMS MSSP specifications. Briefly, in each year, a beneficiary was aligned to an
ACO if the ACO provider organization delivered the largest share of QEM expenditures (relative
to all other provider organizations the beneficiary saw) by primary care practitioners or
specialists with primary care designation (CMS, 2014). Beneficiaries for whom non-ACO
provider organizations delivered the largest share of QEM expenditures constituted the control
group. A list of QEM services is included in Supplementary Table 3.1 and a list of physician and
non-physician specialty codes used in beneficiary assignment are included in Supplementary
Table 3.2. To eliminate potential selection within ACOs and the issue of including alreadytreated groups as controls for later-treated groups, we focused on beneficiaries who were newly
attributed to ACOs or never attributed to ACOs in each year during the intervention period for
the remaining analysis.
We defined each ACO as a collection of physician National Provider Identifiers (NPIs).
Using the ACO Provider-level Research Identifiable Files, we identified physician NPIs
associated with each ACO organization in 2012-2015 and collapsed the QEM expenditures at
beneficiary-provider level to determine whether a beneficiary meets the plurality share of QEM
expenditures to be aligned with ACOs. Consistent with previous research, we used ACO
definitions from the first year of MSSP entry, holding the physician roster constant over the
study period (McWilliams et al., 2016; McWilliams et al., 2020; McWilliams et al., 2018).
Outcomes
The outcome variables were measured using the Cost and Use segment of the Master
Beneficiary Summary Files from 2008 to 2015. Our primary outcome was annual total health

53
care spending which summed up payments made by Medicare, beneficiary, and other primary
payers for all services covered by Medicare Parts A and B. Separately, we calculated total annual
Medicare spending for Parts A and B services. In secondary analyses, we examined annual
health care spending by settings, including inpatient, skilled nursing facility (SNF), hospital
outpatient, physician, home health care, and hospice. We also examined the following measures
of health care utilization by settings: annual number of inpatient stays and length of stay, number
of SNF stays and length of stay, number of emergency room (ER), hospital outpatient, and home
health visits, and number of hospice stays and length of stay. All spending variables were
inflation-adjusted to 2015 dollars using the Personal Consumption Expenditure Price Index for
health care.
Regression Discontinuity Design
To isolate the effects of the ACOs on health care spending and utilization independent of
unobserved differences or biases from selection, we developed comparable treatment and control
groups relying on a fuzzy regression discontinuity design. In the context of the MSSP, the
discontinuity arises from the beneficiary assignment mechanism which requires the plurality
share of QEM expenditures with an ACO provider to be aligned. Based on this, we defined the
treatment assignment variable as the share of QEM expenditures with the ACO minus one
divided by the number of providers the beneficiary visits for QEM services. When this variable
is greater than zero (i.e., share of QEM charges with the ACO is at least as large as it would be if
the charges were split evenly across all providers the beneficiary visits for QEM services), the
beneficiary has a positive probability of ACO alignment.

54
While there is a sharp regression discontinuity among beneficiaries who receive QEM
services from two provider organizations, beneficiaries seeking QEM services from three or
more organizations generate a fuzzier regression discontinuity. For example, consider a
beneficiary with QEM services from three providers organizations: one ACO and two non-ACO
provider organizations. If the share of QEM expenditures with the ACO is 35%, the beneficiary
will be aligned with the ACO only if the share of QEM charges with each non-ACO organization
is less than 35% (e.g., 32.5% each). In other words, ACO alignment is not solely determined by
the strict cutoff rule for beneficiaries seeing three or more providers for QEM services.
The probability of ACO alignment is given by the following equation:
Pr(��!" = 1|� = �) = � + ��!" + �(�)
where � = 1[� ≥ 0] indicates whether the treatment assignment variable exceeds the alignment
cutoff of 0, δ represents the jump in the probability of being aligned at the cutoff point, and �(�)
captures how the probability of alignment increases over the cutoff point. Figure 5 demonstrates
the probability of ACO alignment as a function of the treatment assignment variable. Note that
regardless of the number of providers, beneficiaries with a share of QEM expenditures with the
ACO that was greater than 0.5 were aligned with certainty.
Our regression discontinuity design does not apply to beneficiaries who rely
predominantly on one provider organization for QEM services, and thus they were dropped from
the analysis. This includes beneficiaries who only see one provider group for QEM services
within the year or beneficiaries who see more than one provider but the QEM share with one of
the providers outweighs others (i.e., constitutes over 99% of the beneficiary’s total QEM
expenditures). We compared differences in characteristics between beneficiaries who relied

55
primarily on one provider for QEM services relative to multiple providers. Any differences only
affect interpretation of external validity, not internal validity.
In the context of our study, the regression discontinuity approach assumes that all
determinants of health care spending and utilization for beneficiaries just above or just below the
ACO alignment threshold were similar except for the exposure to MSSP incentives. Therefore,
any differences in outcomes could be interpreted as differences attributable to MSSP ACOs.
Statistical Analysis
We conducted fuzzy regression discontinuity analysis to isolate differences in post-ACO
outcomes between beneficiaries just marginally aligned versus those just marginally unaligned.
For each health care spending and utilization outcome, we estimated the fuzzy regression
discontinuity design by the following two equations:
��!" = � + ��!" + �(�) + �
Y./ = � + ��!" + �(�) + �
where ��!" denotes the ACO alignment status, and Y./ measures health care spending and
utilization outcomes in the years post-ACO alignment. The parameter of interest, �, captures the
effect of ACOs to a beneficiary just marginally aligned. In accordance with fuzzy regression
discontinuity research practices, we estimated the treatment effect using instrumental variables
methods (Lee & Lemieux, 2010; Oldenburg et al., 2016). Specifically, for each outcome, we
fitted beneficiary-level linear regression models to estimate the ACO-related discontinuity in the
relationship between the outcome and the beneficiary’s treatment assignment variable. We
considered outcomes within a relatively narrow bandwidth around the alignment cutoff point,

56
with the treatment assignment variable ranging from -0.1 to 0.1. Standard errors are clustered at
the provider level throughout the analysis.
We estimated both unadjusted and adjusted discontinuities. In the fully adjusted models,
we included the following characteristics of beneficiaries: age, sex, race/ethnicity (White, Black,
Hispanic, Asian, American Indian or Alaska Native, Missing/Other), whether disability was the
original reason for Medicare eligibility, dual enrollment in Medicaid, presence of end-stage renal
disease, and diagnosis with the following chronic conditions: acute myocardial infarction, heart
failure, ischemic heart disease, stroke, diabetes, chronic kidney disease, chronic obstructive
pulmonary disease, and cancer (which combined breast cancer, colorectal cancer, prostate
cancer, lung cancer, and endometrial cancer). Chronic conditions were identified using CMS
Chronic Condition Data Warehouse algorithms directly before the start of MSSP (based on data
from calendar year 2011) and not updated over the study period because ACOs could affect
diagnosis of chronic conditions.
We analyzed discontinuities in outcomes in relation to the 2012 and 2013 ACO
alignment thresholds separately because prior studies suggested that early participants may
respond to MSSP incentives differently from later participants (McWilliams et al., 2016;
McWilliams et al., 2018). Consistent with existing literature, we treated years 2013 to 2015 as
the post-ACO period for both 2012 and 2013 entry cohorts because the 2012 cohort of ACOs did
not enter the MSSP until April or July 2012 (McWilliams et al., 2016; McWilliams et al., 2018).
We conducted several analyses to test the assumptions of our fuzzy regression
discontinuity approach and explore potential sources of bias. First, we checked for evidence of
discontinuities in the distribution of the treatment assignment variable around the ACO
alignment cutoff point. If providers can manipulate beneficiaries to fall just above or below the

57
alignment threshold, then we would expect to see a discontinuity around the alignment cutoff
point. Second, we checked the balance of predetermined beneficiary characteristics and selected
chronic conditions across the alignment threshold in the pre-ACO period to determine whether
alignment thresholds were systematically associated with differences in beneficiary populations.
Third, we further tested for threshold-related discontinuities in health care spending and
utilization outcomes across the alignment threshold in the pre-ACO period.
RESULTS
Beneficiary Characteristics
Table 6 shows sample characteristics at baseline from 2008 to 2012 in relation to the
2013 ACO retrospective alignment threshold. Among 1,634,583 observations who had a QEM
service with an ACO provider, 620,396 (38%) relied mainly on one provider for QEM services
and thus were excluded from our analysis. Among the remaining sample who were eligible for
our regression discontinuity analysis, 85,802 (13.3%) were within 0.1 bandwidth below the
alignment threshold (i.e., control group) and 81,944 (12.7%) were within 0.1 bandwidth above
the alignment threshold (i.e., treatment group). The mean (SD) age was 71.9 (12.11) years
among the control group and 71.6 (12.15) among the treatment group. Females accounted for
63.3% of beneficiaries in the control group and 61.4% of those treated. Differences in the
racial/ethnic composition between the treatment and control groups were mostly small. Among
those treated, 23.4% were disabled and 18.6% were dually eligible for Medicaid. This compares
to 24.0% and 18.8% in the control group, respectively. The treatment group had slightly fewer
number of CCW chronic conditions (6.8 vs. 7.1) and lower rates of ischemic heart disease

58
(48.2% vs. 50.4%), stroke (12.9% vs. 14.4%), chronic obstructive pulmonary disease (24.4% vs.
26.4%), and diabetes (33.6% vs. 35.2%).
The demographic composition and chronic conditions of the one-provider sample were
considerably different from the samples included in regression discontinuity analysis: they
included higher percentages of non-Hispanic Black (7.2%) and Asian (2.0%) persons, lower
percentage of females (57.0%), disabled (20.0%), dual eligible (15.7%), persons with ESRD
(0.6%) and had significantly lower rates of almost all chronic conditions. In Supplementary
Table 3.3, we compared health care spending and utilization across groups at baseline in relation
to the 2013 retrospective alignment threshold. We reported baseline sample characteristics and
health care use and spending separately for the 2012 retrospective alignment threshold and
observed similar patterns (Supplementary Tables 3.4 & 3.5).
Discontinuities in Health Care Spending
Beneficiaries with 2013 alignment scores that exceeded the alignment cutoff point had
significant reductions in post-ACO health care spending outcomes compared to beneficiaries
below the alignment cutoff (Figure 6 and Table 7). Additional graphs demonstrating
discontinuities in health care spending outcomes at the alignment threshold are included in
Supplementary Figure 3.1. Specifically, after adjusting for beneficiary characteristics and
selected chronic conditions, annual total health care spending per beneficiary was $1615 lower
(P<0.001) among beneficiaries above the alignment threshold relative to those below the
threshold in the years post-ACO alignment. This translates to an 8% reduction from the mean
estimate within the bandwidth. Similarly, exceeding the 2013 alignment threshold was associated
with an average reduction of $1396 (or 8%) in annual Medicare spending per beneficiary

59
(P<0.001). ACO alignment eligibility in 2013 was also associated with significantly lower SNF
payment ($305 or 11% reduction, P=0.019), physician payment ($606 or 11% reduction,
P<0.001), and home health payment ($126 or 12% reduction, P=0.01). ACO alignment eligibility
was also associated with a reduction of $409 (or -6%) in inpatient spending although the
discontinuity was barely significant after adjusting for covariates (P=0.13).
Similarly, ACO-related discontinues in health care spending were observed for the 2012
alignment threshold, with larger effect sizes in almost all spending outcomes except for hospice
payments (Table 7). After adjusting for covariates, annual total health care spending and
Medicare spending per beneficiary fell by $2522 (or 12%, P<0.001) and $2186 (or 12%,
P<0.001), respectively, among beneficiaries above the 2012 alignment threshold relative to those
below the threshold. Exceeding the 2012 alignment threshold was also associated with an
adjusted discontinuity of $864 (or 12%) less inpatient payment (P<0.001), $375 (or 13%) less
SNF payment (P<0.001), and $888 (or 14%) less physician payment. Additionally, 2012
alignment threshold was associated with significantly lower spending in hospital outpatient and
home health settings, corresponding to 10% and 11% reductions, respectively. For both 2013 and
2012 cohorts, estimates of adjusted discontinuities for all spending outcomes are within the 95%
confidence intervals of the unadjusted estimates.
Discontinuities in Health Care Utilization
Table 8 shows that ACOs resulted in less intensive use of health care in acuate, postacute, hospital outpatient, and home health care settings for both 2013 and 2012 cohorts. We
reported estimates of unadjusted and adjusted discontinues separately on support of 2013 and
2012 retrospective alignment thresholds. Again, for most health care utilization outcomes,

60
estimates of adjusted discontinuities fell inside the 95% confidence intervals of the unadjusted
estimates.
After adjustment of beneficiary-level characteristics, annual number of inpatient stays
and average length of inpatient stays fell by 13% (P<0.001) and 15% (P<0.001), respectively, for
beneficiaries who were just above the 2013 alignment threshold compared to those just below the
threshold. Being eligible for ACO alignment in 2013 was also associated with significantly less
intense use of SNF care in the years post-ACO alignment, with 20% decrease (P<0.001) in
annual number of SNF stays and 13% decrease (P=0.005) in the average length of SNF stays.
Additionally, ACO alignment in 2013 also resulted in reductions in average counts of ER visits
(-7%, P=0.034), hospital outpatient visits (-8%, P=0.005), and home health visits (-13%,
P=0.014). There were no statistically significant discontinuities in the number of hospice stays
(P=0.971) and days spent in hospice (P=0.224).
A similar pattern was found for the 2012 retrospective alignment threshold (Table 8).
There were meaningful and statistically significant reductions in health care use measures in all
settings with the exception of hospice care. Compared to the 2013 cohort, exceeding the 2012
alignment threshold was associated with greater decrease in several acuate and post-acute care
utilization measures. Specifically, following the 2012 ACO alignment, annual number of
inpatient hospitalizations fell by 0.1 stays per beneficiary (P<0.001) and ER visits fell by 0.2
visits per beneficiary (P<0.001), corresponding to 16% and 17% reductions, respectively. ACO
alignment eligibility in 2012 also resulted in approximately 0.8 less days spent in SNFs
(P<0.001), which translates to 14% reduction from the mean estimate within the bandwidth.
Besides regression discontinuity estimates based on our models, we included scatterplots with

61
superimposed fitted lines to visually demonstrate discontinuities in post-ACO health care
utilization outcomes at the alignment threshold (Supplementary Figure 3.1).
Validating the Discontinuity Design
We conducted several analyses to test the assumptions of our fuzzy regression
discontinuity approach. We presented a histogram of the treatment assignment variable relative
to the alignment threshold value which showed no evidence of clumping around either side of
the threshold (Supplementary Figure 3.2). We found no consistent evidence that predetermined
beneficiary characteristics as well as selected chronic conditions changed discontinuously across
the alignment threshold in the pre-ACO period for both 2013 and 2012 cohorts (Supplementary
Figures 3.3 & 3.4). Additionally, for the most part, health care spending and utilization outcomes
remained continuous across the alignment threshold in the absence of the MSSP incentives for
both 2013 and 2012 cohorts in the pre-ACO period (Supplementary Figures 3.5 & 3.6).
DISCUSSION
In this study, using regression discontinuity design, we found that early cohorts of ACOs
participating in the MSSP have generated large and statistically significant reductions in overall
fee-for-service spending and service utilization for assigned beneficiaries. Three years after the
initial launch of the program, total and Medicare spending fell by 8% among beneficiaries who
were just marginally aligned to ACOs in 2013. Consistent with prior research (McWilliams et
al., 2016; McWilliams et al., 2020; McWilliams et al., 2018), reductions in total and Medicare
spending were even more pronounced, about 12%, among beneficiaries assigned to early ACOs
that entered the program in 2012. Further analysis of fee-for-service expenditures showed that

62
early ACO alignment was associated with significant decreases on several sources of revenues,
including inpatient hospital care, post-acute care in SNFs, hospital outpatient care, physician
care, and home health care. Beneficiaries assigned to early ACOs also showed significant
reductions in the annual number and average length of hospitalizations and SNF stays, as well as
number of visits in ER, hospital outpatient, and home health settings.
Our estimated effect sizes of the MSSP ACOs are significantly larger, both in absolute
term and in percent change, compared to those found in the existing literature (McWilliams et
al., 2016; McWilliams et al., 2020; McWilliams et al., 2018). Several factors could potentially
contribute to the differences. First, prior studies only focused on beneficiaries aligned to ACOs
through primary care services delivered by primary care physicians in office-based settings.
However, in this study, we included beneficiaries assigned to ACOs through all qualified QEM
services in any place of service following CMS MSSP specifications. Moreover, instead of
limiting to services provided by primary care physicians, we followed CMS alignment
methodology in including all qualified primary care practitioners, including additionally
physician assistants, nurse practitioners, and clinical nurse specialists in the beneficiary
alignment process. We also allowed beneficiary alignment to be based on QEM services
provided by certain specialists with primary care designations, should alignment based on QEM
services provided by primary care practitioners fail.
Second, prior studies typically included a very large control group with all beneficiaries
who received the largest share of primary care services from non-ACO providers. However, we
limited our control group to beneficiaries who had at least one QEM service with an ACO
provider because otherwise it’s impossible for them to be attributed to ACOs.

63
Third, we excluded beneficiaries who received QEM services from only one provider in
our regression discontinuity analysis. Such individuals constituted about 40% of our samples and
they had significantly lower health care utilization and spending at baseline compared to those
seeing multiple providers (Supplementary Tables 3.3 & 3.5). Inclusion of these low-use
beneficiaries could dilute the estimates of the impact of ACOs since there could be very little
room for ACOs to cut wasteful spending and improve care coordination across providers if
beneficiaries only rely on one provider for primary care services.
Finally, the impact we estimated in regression discontinuity design was a local average
treatment effect that is relevant to a specific group of beneficiaries: individuals who seek care
relatively evenly among different providers. Therefore, this effect could be different from the
average treatment effects across all individuals estimated in prior studies.
Our study has several limitations. Rather than defining ACOs as a collection of taxpayer
identification numbers, we defined ACOs as a collection of physician NPIs. However, as showed
in previous studies, alternative definition of ACOs had minimal impact on outcomes
(McWilliams et al., 2016; McWilliams et al., 2020; McWilliams et al., 2018). Second, our
estimates were based on early cohorts of ACOs and may not be directly generalizable to ACOs
entering the MSSP in more recent years. Third, similar to previous studies, our results may still
underestimate the impact of ACOs because of potential spillover effects. For example, this
includes more efficient health care delivery for beneficiaries not aligned to ACOs. Finally, we
focused on the impact of ACOs on health care utilization and spending. Quality of care was not
directly examined in the current study.
Despite the limitations, findings from this study show that MSSP ACOs have been
considerably more successful at reducing spending and utilization than previously understood,

64
particularly among patients who stand to gain the most from coordinated care efforts. Our results
have important implications for policymakers interested in re-designing health care payment and
insurance models.

65
TABLES & FIGURES
Figure 5. Probability of ACO Alignment as a Function of the Alignment Threshold
Notes: Probability of ACO alignment as a function of the alignment threshold from pooled 2012 and 2013 samples.
The treatment assignment variable is defined as the share of QEM expenditures with the ACO provider minus one
divided by the number of providers. For example, a beneficiary with two providers and share QEM expenditures less
than 0.5 will have a negative forcing variable and has zero probability of alignment; that beneficiary with two
providers and share QEM expenditures greater than 0.5 will have a positive forcing variable and will be aligned with
certainty. Note that regardless of the number of providers, beneficiaries with share QEM expenditures greater than
0.5 are aligned with certainty.
0 .2 .4 .6 .8 1
Probability of ACO-Alignment
-.6 -.4 -.2 0 .2 .4 .6 .8 1
Treatment Assignment Variable

66
Table 6. Sample Characteristics at Baseline in Relation to the 2013 ACO Retrospective
Alignment Threshold
One Provider Control Treatment
N, bene-years 620,396 85,802 81,944
Mean Age (SD) 72.02 (11.46) 71.89 (12.11) 71.64 (12.15)
<65, % 14.81 17.23 17.13
65-69, % 17.61 16.03 16.15
70-74, % 24.59 22.10 23.07
75-79, % 19.02 18.52 18.83
80-84, % 13.76 14.72 14.05
85+, % 10.21 11.41 10.78
Female, % 57.01 63.27 61.42
Race/Ethnicity, %
White 86.51 87.44 87.50
Black 7.18 6.53 6.49
Hispanic 3.33 3.57 3.51
Asian 1.96 1.41 1.54
American Indian/Alaska Native 0.14 0.34 0.27
Missing/Other 0.88 0.70 0.70
Disabled, % 19.95 23.95 23.43
Dual, % 15.71 18.83 18.63
End-Stage Renal Disease, % 0.57 0.77 0.85
CCW Chronic Conditions, No. 6.01 7.09 6.81
Chronic Conditions, %
AMI 3.68 3.98 3.99
Heart Failure 18.42 24.69 24.18
Ischemic Heart Disease 42.28 50.36 48.19
Stroke 10.43 14.36 12.89
COPD 19.13 26.36 24.42
Diabetes 31.00 35.21 33.62
Chronic Kidney Disease 15.48 20.47 19.86
Cancer 13.41 14.87 14.30
Notes: Data from 2008-2012. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded from regression
discontinuity analysis. Beneficiaries with a treatment assignment variable between -0.1 and 0 constitutes the control
group while beneficiaries with a treatment assignment variable between 0 and 0.1 are the treatment group.

67
5000 6000 7000 8000 -.2 -.1 0 .1 .2
14000 15000 16000 17000 18000 19000 -.2 -.1 0 .1 .2 1500 2000 2500 3000 -.2 -.1 0 .1 .2
2500 3000 3500 4000 -.2 -.1 0 .1 .2
5000 5500 6000 6500 -.2 -.1 0 .1 .2
16000 18000 20000 22000 -.2 -.1 0 .1 .2
Figure 6. MSSP ACO-Related Discontinuities in Annual Health Care Spending, 2013 ACO
Retrospective Alignment Threshold
Total Payment Per Beneficiary
Medicare Payment Per Beneficiary
Inpatient Payment Per Beneficiary
Skilled Nursing Facility Payment Per
Beneficiary
Hospital Outpatient Payment Per Beneficiary
Physician Payment Per Beneficiary
ACO’s QEM Share – (1/Number of Providers)
Notes: Data from 2013 to 2015. This figure presents scatterplots of health care payments with evenly spaced bins on
the support of the 2013 ACO retrospective alignment threshold. The horizontal axis is the treatment assignment
variable which is defined as the share of QEM expenditures with ACO provider minus one divided by the number of
providers. Each point in the graphs is an average calculated among all beneficiaries within a binned treatment
assignment variable and is adjusted for beneficiary characteristics and chronic conditions. Fitted values from
regression discontinuity models, adjusted for beneficiary characteristics and chronic conditions are superimposed on
the scatterplots (red lines).

68
Table 7. Discontinuities in Health Care Spending in relation to 2013 and 2012 ACO Retrospective Alignment Thresholds
Measure Unadjusted
Discontinuity (95% CI) P Value Adjusted Discontinuity
(95% CI) P Value Mean in
Bandwidth
Percent
Change
2013 Alignment Threshold
Total health care payment -1956 (-3113 to -798) <0.001 -1615 (-2609 to -621) <0.001 19900 -8%
Medicare payment -1696 (-2681 to -710) <0.001 -1396 (-2237 to -555) <0.001 16768 -8%
Inpatient payment -499 (-1086 to 87.19) 0.095 -409 (-939 to 121) 0.13 6639 -6%
SNF payment -411 (-685 to -138) 0.003 -305 (-559 to -51) 0.019 2698 -11%
Hospital outpatient payment -147 (-426 to 133) 0.302 -115 (-353 to 122) 0.341 3215 -4%
Physician payment -658(-935 to -380) <0.001 -606 (-873 to -339) <0.001 5880 -11%
Home health payment -163 (-269 to -58) 0.002 -126 (-223 to -30) 0.01 1091 -12%
Hospice payment -77 (-181 to 27) 0.148 -54 (-158 to 50) 0.311 377 -14%
2012 Alignment Threshold
Total health care payment -3383 (-3957 to -2809) <0.001 -2522 (-3013 to -1705) <0.001 20793 -12%
Medicare payment -3381 (-4905 to -1857) <0.001 -2186 (-2610 to -1317) <0.001 17539 -12%
Inpatient payment -1180 (-1461 to -899) <0.001 -846 (-1098 to -595) <0.001 6845 -12%
SNF payment -512 (-666 to -359) <0.001 -375 (-515 to -234) <0.001 2821 -13%
Hospital outpatient payment -424 (-566 to -283) <0.001 -314 (-435 to -195) <0.001 3229 -10%
Physician payment -1079 (-1269 to -889) <0.001 -888 (-1069 to -707) <0.001 6248 -14%
Home health payment -193 (-255 to -132) <0.001 -127 (-181 to -73) <0.001 1198 -11%
Hospice payment 5 (-69 to 80) 0.886 28 (-46 to 102) 0.454 453 6%
Notes: Data from 2013 to 2015. Sample is beneficiaries with more than one QEM provider retrospectively in 2013 and 2012. Fuzzy regression discontinuity
estimates are shown, with common bandwidth of 0.1 on both sides of the alignment threshold. Mean in bandwidth refers to mean estimate calculated within the
bandwidth of 0.1 around the cutoff. Total health care payments sum up payments by Medicare, beneficiary, and other primary payer for services covered by Parts
A and B. Medicare payments refer to payments by Medicare for Parts A and B services. Physician payments sum up eleven cost/use categories from Carrier files.
Payments were converted to 2015 dollars. The 95% confidence intervals were estimated using robust standard errors clustered at the provider level.

69
Table 8. Discontinuities in Health Care Utilization in relation to 2013 and 2012 ACO Retrospective Alignment Thresholds
Measure Unadjusted
Discontinuity (95% CI) P Value Adjusted Discontinuity
(95% CI) P Value Mean in
Bandwidth
Percent
Change
2013 Alignment Threshold
Inpatient stays -0.08 (-0.11 to -0.04) <0.001 -0.07 (-0.10 to -0.04) <0.001 0.53 -13%
Inpatient length of stay, days -0.50 (-0.80 to -0.20) <0.001 -0.45 (-0.72 to -0.18) <0.001 3.06 -15%
SNF stays -0.04 (-0.06 to -0.03) <0.001 -0.04 (-0.05 to -0.02) <0.001 0.20 -20%
SNF length of stay, days -0.91 (-1.44 to -0.38) <0.001 -0.70 (-1.19 to -0.21) 0.005 5.24 -13%
ER visits -0.09 (-0.17 to -0.01) 0.033 -0.08 (-0.16 to -0.01) 0.034 1.13 -7%
Hospital outpatient visits -1.16 (-2.03 to -0.29) 0.009 -0.92 (-1.56 to -0.29) 0.005 11.52 -8%
Home health visits -1.04 (-1.75 to -0.34) 0.004 -0.82 (-1.47 to -0.17) 0.014 6.44 -13%
Hospice stays -0.001 (-0.006 to 0.003) 0.581 -0.000 (-0.005 to 0.005) 0.971 0.03 0%
Hospice length of stay, days -0.55 (-1.22 to 0.11) 0.103 -0.41 (-1.08 to 0.25) 0.224 2.30 -18%
2012 Alignment Threshold
Inpatient stays -0.12 (-0.13 to -0.10) <0.001 -0.09 (-0.10 to -0.06) <0.001 0.55 -16%
Inpatient length of stay, days -0.61 (-0.75 to -0.32) 0.001 -0.43 (-0.56 to -0.29) <0.001 3.26 -13%
SNF stays -0.04 (-0.05 to -0.02) <0.001 -0.03 (-0.04 to -0.02) <0.001 0.20 -15%
SNF length of stay, days -1.03 (-1.33 to -0.73) <0.001 -0.76 (-1.04 to -0.48) <0.001 5.45 -14%
ER visits -0.26 (-0.30 to -0.21) <0.001 -0.20 (-0.24 to -0.16) <0.001 1.15 -17%
Hospital outpatient visits -1.42 (-1.86 to -0.97) 0.001 -0.97 (-1.31 to -0.63) <0.001 11.73 -8%
Home health visits -1.36 (-2.58 to -0.13) 0.030 -0.66 (-1.04 to -0.28) 0.001 7.21 -9%
Hospice stays -0.001 (-0.004 to 0.002) 0.668 -0.000 (-0.002 to 0.004) 0.627 0.03 0%
Hospice length of stay, days 0.07 (-0.41 to 0.54) 0.783 0.20 (-0.26 to 0.67) 0.392 2.77 7%
Notes: Data from 2013 to 2015. Sample is beneficiaries with more than one QEM provider retrospectively in 2013 and 2012. Fuzzy regression discontinuity
estimates are shown, with common bandwidth of 0.1 on both sides of the alignment threshold. Mean in bandwidth refers to mean estimate calculated within the
bandwidth of 0.1 around the cutoff. The 95% confidence intervals were estimated using robust standard errors clustered at the provider level.

70
APPENDIX
Supplementary Table 3.1. Qualified Evaluation and Management Services
Description HCPCS Codes
Office and other outpatient services
New patient visit 99201-99205
Established patient visit 99211-99215
Nursing facility services
Initial new or established patient visit 99304-99306
Subsequent new or established patient visit 99307-99310
Discharge new or established patient services 99315-99316
Other new or established patient service 99318
Domiciliary, rest home, or custodial care services
New patient visit 99324-99328
Established patient visit 99334-99337
Oversight services 99339-99340
Home services
New patient visit 99341-99345
Established patient visit 99347-99350
Chronic care management service 99490
Transitional care management services 99495-99496
Wellness visits
Welcome visit G0402
Annual wellness visit G0438-G0439
Hospital outpatient clinic visit G0463
Note: The HCPCS codes are taken from the CMS Medicare Shared Savings Program Beneficiary Assignment
documentation.

71
Supplementary Table 3.2. Physician and Non-Physician Specialty Codes Used in
Beneficiary Assignment
Description Specialty Code
Primary care practitioner
General practice 01
Family practice 08
Internal medicine 11
Geriatric medicine 38
Nurse practitioner 50
Clinical nurse specialist 89
Physician assistant 97
Specialist with primary care designation
Cardiology 06
Osteopathic manipulative medicine 12
Neurology 13
Obstetrics/gynecology 16
Sports medicine 23
Physical medicine and rehabilitation 25
Psychiatry 26
Geriatric psychiatry 27
Pulmonary disease 29
Nephrology 39
Endocrinology 46
Multispecialty clinic or group practice 70
Addiction medicine 79
Hematology 82
Hematology/oncology 83
Preventive medicine 84
Neuropsychiatry 86
Medical oncology 90
Gynecologist/oncologist 98
Note: The specialty codes are taken from the CMS Medicare Shared Savings Program Beneficiary Assignment
documentation. Primary care practitioners (including primary care physicians and non-physician specialty codes)
were used in claims-based assignment Step 1. Specialists with primary care designation were only used in
assignment Step 2 to beneficiaries without primary care services rendered by any primary care practitioners either
inside or outside ACOs.

72
Supplementary Table 3.3. Health Care Spending and Utilization at Baseline in Relation to
the 2013 ACO Retrospective Alignment Threshold
One Provider Control Treatment
N, bene-years 620,396 85,802 81,944
Total payment 7744 12048 11348
Inpatient payment 2323 3750 3588
Skilled nursing facility payment 390 852 767
Hospital outpatient payment 1551 2147 2086
Physician payment 3160 4721 4383
Home health payment 296 564 501
Hospice payment 25 16 22
Medicare payment 6415 10102 9504
Inpatient Medicare payment 2109 3413 3267
Skilled nursing facility Medicare payment 339 726 653
Hospital outpatient Medicare payment 1199 1676 1627
Physician Medicare payment 2447 3707 3433
Home health Medicare payment 295 563 501
Hospice Medicare payment 25 16 22
Inpatient stays 0.21 0.35 0.32
Inpatient days 0.95 1.72 1.58
Emergency room visits 0.44 0.77 0.70
Skilled nursing facility stays 0.03 0.06 0.06
Skilled nursing facility days 0.73 1.63 1.47
Hospital outpatient visits 5.93 8.09 7.94
Home health visits 1.71 3.29 2.89
Hospice stays 0.001 0.001 0.001
Hospice days 0.15 0.09 0.14
Notes: Data from 2008-2012. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded from regression
discontinuity analysis. Beneficiaries with a treatment assignment variable between -0.1 and 0 constitutes the control
group while beneficiaries with a treatment assignment variable between 0 and 0.1 are the treatment group. All
spending variables were inflation-adjusted to 2015 dollars.

73
Supplementary Table 3.4. Sample Characteristics at Baseline in Relation to the 2012 ACO
Retrospective Alignment Threshold
One Provider Control Treatment
N, bene-years 504,094 68,284 64,071
Mean Age (SD) 71.83 (11.15) 71.37 (12.29) 71.47 (12.21)
<65, % 14.31 18.29 17.40
65-69, % 19.99 16.92 17.72
70-74, % 24.20 21.39 21.84
75-79, % 18.83 18.40 18.13
80-84, % 13.41 14.57 14.35
85+, % 9.26 10.43 10.57
Female, % 57.48 63.99 62.18
Race/Ethnicity, %
White 86.00 86.35 86.72
Black 7.13 7.41 7.12
Hispanic 3.80 3.60 3.51
Asian 2.11 1.63 1.73
American Indian/Alaska Native 0.16 0.29 0.25
Missing/Other 0.81 0.72 0.67
Disabled, % 18.66 24.93 23.26
Dual, % 15.07 20.62 19.14
End-Stage Renal Disease, % 0.57 0.82 0.92
CCW Chronic Conditions, No. 5.89 7.17 6.85
Chronic Conditions, %
AMI 3.41 4.04 3.84
Heart Failure 18.32 26.45 24.81
Ischemic Heart Disease 42.49 52.67 50.49
Stroke 10.34 14.53 13.49
COPD 18.37 27.27 24.89
Diabetes 31.13 37.07 34.95
Chronic Kidney Disease 14.30 20.30 19.10
Cancer 13.30 14.72 14.71
Notes: Data from 2008-2011. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded from regression
discontinuity analysis. Beneficiaries with a treatment assignment variable between -0.1 and 0 constitutes the control
group while beneficiaries with a treatment assignment variable between 0 and 0.1 are the treatment group.

74
Supplementary Table 3.5. Health Care Spending and Utilization at Baseline in Relation to
the 2012 ACO Retrospective Alignment Threshold
One Provider Control Treatment
N, bene-years 504,094 68,284 64,071
Total payment 7630 12812 11725
Inpatient payment 2322 4024 3628
Skilled nursing facility payment 369 902 793
Hospital outpatient payment 1389 2136 2038
Physician payment 3251 5130 4696
Home health payment 282 605 548
Hospice payment 15 14 22
Medicare payment 6349 10794 9859
Inpatient Medicare payment 2130 3698 3333
Skilled nursing facility Medicare payment 322 768 678
Hospital outpatient Medicare payment 1075 1670 1589
Physician Medicare payment 2524 4039 3689
Home health Medicare payment 282 605 548
Hospice Medicare payment 15 14 22
Inpatient stays 0.21 0.38 0.33
Inpatient days 0.97 1.88 1.68
Emergency room visits 0.40 0.06 0.06
Skilled nursing facility stays 0.03 1.71 1.47
Skilled nursing facility days 0.68 0.81 0.69
Hospital outpatient visits 5.28 8.17 7.71
Home health visits 1.56 3.42 3.08
Hospice stays 0.000 0.000 0.000
Hospice days 0.10 0.09 0.13
Notes: Data from 2008-2011. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded from regression
discontinuity analysis. Beneficiaries with a treatment assignment variable between -0.1 and 0 constitutes the control
group while beneficiaries with a treatment assignment variable between 0 and 0.1 are the treatment group. All
spending variables were inflation-adjusted to 2015 dollars.

75
.4 .45 .5 .55 .6 -.2 -.1 0 .1 .2
200 300 400 500 600 -.2 -.1 0 .1 .2
2 2.5 3 3.5 -.2 -.1 0 .1 .2
.14 .16 .18 .2 .22 .24 -.2 -.1 0 .1 .2
3.5 4 4.5 5 5.5 6 -.2 -.1 0 .1 .2
900 1000 1100 1200 1300 -.2 -.1 0 .1 .2
Supplementary Figure 3.1. MSSP ACO-Related Discontinuities in Annual Health Care
Spending and Utilization, 2013 ACO Retrospective Alignment Threshold
Home Health Payment Per Beneficiary
Hospice Payment Per Beneficiary
No. Inpatient Stays Per Beneficiary
Inpatient Days Per Beneficiary
No. SNF Stays Per Beneficiary
SNF Days Per Beneficiary
ACO’s QEM Share – (1/Number of Providers)

76
5 6 7 8 -.2 -.1 0 .1 .2
10.5 11 11.5 12 12.5 13 -.2 -.1 0 .1 .2 .02 .025 .03 .035 .04 .045 -.2 -.1 0 .1 .2
1 2 3 4 -.2 -.1 0 .1 .2 .9 1 1.1 1.2 1.3 1.4 -.2 -.1 0 .1 .2
No. ER Visits Per Beneficiary
No. Hospital Outpatient Visits Per Beneficiary
No. Home Health Visits Per Beneficiary
No. Hospice Stays Per Beneficiary
Hospice Days Per Beneficiary
ACO’s QEM Share – (1/Number of Providers)
Notes: Data from 2013 to 2015. This figure presents scatterplots of health care payments with evenly spaced bins on
the support of the 2013 ACO retrospective alignment threshold. Each point in the graphs is an average calculated
among all beneficiaries within a binned forcing variable, share of QEM expenditures with ACO provider minus one
divided by the number of providers (indicated on the horizontal axis) and is adjusted for beneficiary characteristics
and indicators of chronic conditions. Fitted values from regression discontinuity models, adjusted for beneficiary
characteristics and chronic conditions are superimposed on the scatterplots (red lines). The vertical distance between
the fitted lines at the alignment threshold (dashed vertical line) correspond to the regression discontinuity estimates.

77
Supplementary Figure 3.2. Histogram of the Alignment Threshold Score
Notes: Density of the forcing variable from 2012-2015. Densities are reported within bins of width 0.02. Sample
consists of beneficiaries who have seen an ACO provider and received QEM services from more than one provider.
0 1 2 3 4
Percent
-.6 -.4 -.2 0 .2 .4 .6 .8 1
Treatment Assignment Variable

78
Supplementary Figure 3.3. Patient Characteristics at Baseline by Position Relative to the
2013 Retrospective Alignment Threshold
Age Age less than 65
Female White
Black Hispanic
Asian AIAN .82 .84 .86 .88 .9 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.05 .06 .07 .08 .09 .1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.02 .04 .06 .08 .1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .005 .01 .015 .02 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .005 .01 .015 .02 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

79
Dual Eligible Disabled
ESRD Cancer
Acute Myocardial Infraction Stroke
COPD Heart Failure

80
Ischemic Heart Disease Diabetes
Chronic Kidney Disease
Notes: Data from 2008-2012. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded. We showed the
integrated means squared error (IMSE) optimal number of bins for evenly spaced bins on the support of the 2013
retrospective alignment threshold (share of QEM expenditures minus one divided by the number of providers). The
bandwidths are chosen to span the full support of the data.

81
Supplementary Figure 3.4. Patient Characteristics at Baseline by Position Relative to the
2012 Retrospective Alignment Threshold
Age Age less than 65
Female White
Black Hispanic
Asian AIAN
70 71 72 73 74 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.14 .15 .16 .17 .18 .19 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.62 .64 .66 .68 .7 .72 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.75 .8 .85 .9 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.06 .08 .1 .12 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.02 .03 .04 .05 .06 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.01 .015 .02 .025 .03 .035 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .002 .004 .006 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

82
Dual Eligible Disabled
ESRD Cancer
Acute Myocardial Infraction Stroke
COPD Heart Failure
.15 .2 .25 .3 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.2 .22 .24 .26 .28 .3 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .005 .01 .015 .02 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.05 .1 .15 .2 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.005 .01 .015 .02 .025 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.02 .04 .06 .08 .1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.1 .15 .2 .25 .3 .35 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.1 .2 .3 .4 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

83
Ischemic Heart Disease Diabetes
Chronic Kidney Disease
Notes: Data from 2008-2011. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded. We showed the
integrated means squared error (IMSE) optimal number of bins for evenly spaced bins on the support of the 2012
retrospective alignment threshold (share of QEM expenditures minus one divided by the number of providers). The
bandwidths are chosen to span the full support of the data. .2 .4 .6 .8 1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.3 .4 .5 .6 .7 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.1 .2 .3 .4 .5 .6 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

84
Supplementary Figure 3.5. Pre-ACO Outcomes by Position Relative to the 2013
Retrospective Alignment Threshold
Total Health Care Payment Total Medicare Payment
Inpatient Payment SNF Payment
Hospital Outpatient Payment Physician Payment
10000 15000 20000 25000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
8000 10000 12000 14000 16000 18000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
2000 3000 4000 5000 6000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
500 1000 1500 2000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1500 2000 2500 3000 3500 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
4000 6000 8000 10000 12000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

85
0 .0005 .001 .0015 .002
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Home Health Payment Hospice Payment Regression function fit
Inpatient Stays Inpatient Length of Stay
SNF Stays SNF Length of Stay
400 600 800 1000 1200
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.2 .3 .4 .5 .6
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1 1.5 2 2.5 3 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.04 .06 .08 .1 .12
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1 2 3 4 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

86
ER Visits Hospital Outpatient Visits
Home Health Visits Hospice Stays
Hospice Length of Stay
Notes: Data from 2008-2012. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded. We showed the
integrated means squared error (IMSE) optimal number of bins for evenly spaced bins on the support of the 2013
retrospective alignment threshold (share of QEM expenditures minus one divided by the number of providers). The
bandwidths are chosen to span the full support of the data. Payments were converted to 2015 dollars.
0 1 2 3 4 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
8 10 12 14 16
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 2 4 6 8 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .0005 .001 .0015 .002
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .1 .2 .3 .4
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

87
Supplementary Figure 3.6. Pre-ACO Outcomes by Position Relative to the 2012
Retrospective Alignment Threshold
Total Health Care Payment Total Medicare Payment
Inpatient Payment SNF Payment
Hospital Outpatient Payment Physician Payment
10000 15000 20000 25000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
5000 10000 15000 20000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
3000 4000 5000 6000 7000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
600 800 1000 1200 1400 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1500 2000 2500 3000 3500 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
4000 6000 8000 10000 12000 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

88
Home Health Payment Hospice Payment
Inpatient Stays Inpatient Length of Stay
SNF Stays SNF Length of Stay
400 500 600 700 800 900
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 50 100 150
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.3 .4 .5 .6 .7 .8
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1 1.5 2 2.5 3 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.04 .06 .08 .1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
1 1.5 2 2.5
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

89
ER Visits Hospital Outpatient Visits
Home Health Visits Hospice Stays
Hospice Length of Stay
Notes: Data from 2008-2011. Sample includes beneficiaries continuously enrolled in Medicare parts A and B and
who saw an ACO provider for QEM services within the year. Beneficiaries relied on one provider for QEM services
(those saw only one provider or QEM share with one provider is larger than 0.999) are excluded. We showed the
integrated means squared error (IMSE) optimal number of bins for evenly spaced bins on the support of the 2012
retrospective alignment threshold (share of QEM expenditures minus one divided by the number of providers). The
bandwidths are chosen to span the full support of the data. Payments were converted to 2015 dollars. .6 .8 1 1.2 1.4
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
6 8 10 12
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
2 3 4 5 6 7 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
.0005 .001 .0015 .002 .0025
-.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit
0 .2 .4 .6 .8 1 -.5 0 .5 1
Sample average within bin Polynomial fit of order 4
Regression function fit

90
Chapter 4: Association of Dementia Severity at Diagnosis with Health Care Utilization and
Costs around the Time of Incident Diagnosis
INTRODUCTION
Dementia affects 7.6 million older U.S. adults (Alzheimer’s Association, 2023) and
exerts a heavy economic burden on society. The average annual total cost for a person aged 70
and older living with dementia were over $70,000 in 2010 and are projected to double by 2050;
medical costs accounted for 60% of the annual total cost in 2010 and are about to rise to 76% in
2050 (Zissimopoulos et al., 2014). Diagnosis at more advanced dementia stages is associated
with higher health care utilization and costs and creates additional economic strain on families
and the health system due to increased preventable hospitalizations, emergency department visits
and frequent transitions of care (Anderson et al., 2020; Barnett et al., 2014; Callahan et al., 2012;
Gozalo et al., 2011; Knox et al., 2020; LaMantia, Lane, et al., 2016; LaMantia, Stump, et al.,
2016; Wolf et al., 2019; Carolyn W. Zhu et al., 2015). Early diagnosis of dementia, before
symptoms impair cognitive and physical processes, may be an effective venue for reducing
medical costs. Early diagnosis also supports opportunities for clinical trial participation and
eligibility for new dementia treatments (van Dyck et al., 2023), and facilitates connecting to
social services and planning for future financial, health care and long-term care needs (Dubois et
al., 2016; Robinson et al., 2015). Despite these important benefits of early dementia diagnosis,
studies have reported delayed dementia diagnoses in the older U.S. population with higher rates
among Black and Hispanic persons compared to non-Hispanic white persons (Amjad et al.,
2018; Chen et al., 2019; Lin et al., 2021).

91
There is considerable evidence that persons with dementia incur higher medical costs
than dementia-free persons (Coe et al., 2023; Hurd et al., 2013; Hurd et al., 2015; Lin et al.,
2016; White et al., 2019; C. W. Zhu et al., 2015). Substantial increases in medical costs are
observed around the time of dementia diagnosis (Lin et al., 2016; White et al., 2019; C. W. Zhu
et al., 2015) that are associated with intensive use of acute and post-acute services (Bynum et al.,
2004; Coe et al., 2023; Daras et al., 2017; Desai et al., 2019; Hoffman et al., 2022; Carolyn W.
Zhu et al., 2015; C. W. Zhu et al., 2015). However, these prior estimates reflect averages of
persons in various stages of disease progression, and persons diagnosed at different stages of
dementia will utilize health care differently. Data limitations have generally precluded
population-based estimates of differences in health care use and costs around the time of
diagnosis for persons diagnosed at different stages or severity of dementia thus limiting
understanding of how improving detection for early diagnosis may impact health care use and
costs for society.
In this study, we bring together rich population survey data from the Health and
Retirement Study (HRS), the Aging, Demographics, and Memory Study (ADAMS), and
Medicare claims administrative data to quantify health care use and costs around the time of
incident dementia diagnosis for a population representative sample of three groups of older
adults: those diagnosed at early/mild stage disease; those diagnosed at moderate stage disease;
and those diagnosed at severe stage of disease. We use data on respondents aged 70 and older
from the HRS and their Medicare claims and analyze inpatient, emergency room, and outpatient
health care use and associated costs in the quarter of their first diagnosis and for 4 quarters
before and after the diagnosis. We quantify health care use and costs for persons with different
levels of dementia severity at the time of diagnosis adjusting for socio-economic, demographic

92
and health differences. We build on prior study documenting changes in health care use and
spending around time of diagnosis while adding heterogeneity for persons diagnosed at different
severity of disease for insight on health care use and cost implications of early detection in the
older U.S. population.
METHODS
Data and Study Population
This study used data from the 2000-2016 waves of HRS with data linkage to Medicare
claims for respondents enrolled in traditional Medicare (TM). The HRS is a longitudinal study
that has surveyed nationally representative samples of adults older than 50 years in the U.S. and
their spouses every two years since 1992. The HRS collects information on socio-demographic
characteristics, health status, health care use and associated costs, employment, and financial
resources of respondents and their spouses (Juster & Suzman, 1995). Eighty-eight percent of
HRS respondents opted to link their HRS survey data to Medicare claims (St Clair et al., 2017).
Medicare claims data is an important data source for studying dementia diagnosis and health care
utilization and spending in the older U.S. population as it includes information on all diagnoses,
health care use and their costs for almost all (97%) U.S. adults older than 65 and is not subject to
recall bias of survey reports.
This study also used the ADAMS, a nationally representative study of dementia with a
subsample of HRS respondents aged 70 years and above (Langa et al., 2005). Because the
ADAMS performs thorough assessment of dementia, including a validated measure of dementia
severity rated by a clinical professional, we used data from the ADAMS to model dementia
severity and apply model parameters to the longitudinal HRS. We describe our methods below

93
and supplemental documentation provides details and code for replication of the severity
measure.
We identified a retrospective cohort of older adults with incident dementia from HRS
respondents with linked TM claims in the years 2000-2016. Incident dementia was identified in
TM claims using ICD-9 and ICD-10 codes listed in Supplementary Table 4.1 and a rigorous
algorithm for measuring diagnosed dementia in claims that is described in detail in a prior
publication and publicly available (Thunell et al., 2019). Briefly, dementia cases were
ascertained using a combination of dementia diagnosis codes and dementia symptom codes. To
ensure that we capture incident dementia, we required a two-year ‘wash-out’ period with no
dementia diagnosis prior to the year of incident dementia diagnosis. To exclude potential rule-out
diagnosis, we required a dementia diagnosis to be followed by a second diagnosis or symptom
code within two years or death within one year. Dementia diagnosis and symptom codes are
identified in TM claims from the following settings: inpatient, outpatient, skilled nursing facility,
home health care, and carrier claims. Dementia symptom codes are used at a different time point
for verification purposes and are only used together with dementia diagnosis codes for
identifying dementia cases. Previous study finds that this algorithm improves identification of
dementia cases among racial minority populations in claims data (Thunell et al., 2019).
We restricted the sample to individuals aged 70 years or older at incident dementia to
match the ADAMS sample. To reduce measurement error of dementia severity at time of
diagnosis, we imposed an asymmetrical time window restriction which required the matched
HRS interview to be up to 12 months before or up to 6 months after incident dementia that
occurred no earlier than 2000 (Chen et al., 2019). The sample selection criteria are detailed in
Table 9. The final study sample of 2015 older adults with incident dementia consisted of 1577

94
non-Hispanic Whites (78.3%), 285 non-Hispanic Blacks (14.1%), 119 Hispanics (5.9%), and 34
other races (1.7%).
Dementia Severity at Diagnosis
We modeled dementia severity using the validated Clinical Dementia Rating Scale
(CDR). The CDR is a widely used five-point dementia staging tool that clinically measures
dementia severity based on individual’s performance in six areas: memory, orientation, judgment
and problem-solving, community affairs, home and hobbies, and personal care (Langa et al.,
2005). In the ADAMS, a trained clinical professional determines the CDR score for each subject
based on information collected from a structured interview with both subjects and their
informants (Morris, 1993). We used Poisson regressions to model the CDR score based on data
from 852 ADAMS respondents using variables also available to all HRS respondents, including
age, sex, race, education level, cognitive function, activities of daily living (ADLs), instrumental
activities of daily living (IADLs), depression status, and whether a proxy responded for the
subject. Supplementary Table 4.2 provides estimates from the Poisson regression.
Supplementary Figure 4.1 shows the density curve of the full distribution of the predicted CDR
scores as compared to the observed CDR scores in the ADAMS. In Supplementary Table 4.3, we
include additional assessment of the within-sample model fit by comparing the distribution of
predicted CDR categories (predicted CDR scores are rounded to the 0.25 place and binned) with
the distribution of observed CDR categories in ADAMS.
We applied the model parameters from Poisson regressions estimated in ADAMS to all
HRS respondents older than 70 years with incident dementia in TM claims. We predicted CDR
using data from the closest interview to the time of incident diagnosis for each HRS respondent

95
diagnosed with dementia. Based on the predicted CDR score at diagnosis, we classified
individuals into three categories of dementia severity: mild dementia (CDR<0.5); moderate
dementia (0.5≤CDR<2.5); and severe dementia (CDR≥2.5). Additional Technical
Documentation provides details for recreating the dementia severity measure including how the
variables were operationalized and recoded, and how missing observations were handled and
results from tests of sensitivity of the predicted estimates to the Poisson model.
Health Care Utilization and Spending
The outcome variables are measured using Medicare claims data. These data contain
information about service use and payments by Medicare, beneficiary, and other payers. We
measured quarterly health care utilization and spending in different health care settings:
inpatient, emergency room (ER), and outpatient. We used inpatient claim files to identify
inpatient stays, length of stay, and expenditures. ER visits were identified in inpatient and
hospital outpatient claims files based on established revenue center codes (0450-0459, 0981).
Hospital outpatient and carrier (physician) files were used to obtain outpatient visits and
expenditures. We calculated quarterly health care utilization and expenditures in the four
quarters before, the quarter of, and the four quarters after the date of incident dementia diagnosis.
For health care spending, we summed up payments made by Medicare, beneficiary, and other
primary payers to get total expenditures. Expenditures were inflation-adjusted to 2016 dollars
using the Personal Consumption Expenditure Price Index for health care.
Statistical Analysis

96
We report descriptive statistics of the study population for each of the three dementia
severity categories at diagnosis and overall. Next, we used an event study framework to model
health care use and costs adjusted for observable differences that may impact use and spending
for persons independent of dementia severity at diagnosis. To compare health care utilization and
spending across severity groups over time, we predicted health care use and spending outcomes
at each time point with other covariates set at overall sample means and report model predictions
of use and costs for each category of dementia severity.
The models include interactions between dementia severity at diagnosis (mild, moderate,
severe) and quarters relative to the date of incident diagnosis (-4, -3, -2, …, +2, +3, +4, with
quarter 0 indicating the quarter of incident dementia diagnosis) to study the association of
dementia severity at diagnosis on health care utilization and spending before, at and after the
diagnosis. The interaction term between quarter -4 and mild severity was excluded such that the
other indicators captured changes relative to this initial quarter for individuals diagnosed at mild
stage of dementia. Individuals deceased after dementia diagnosis were dropped from subsequent
quarters after the quarter of death in the analysis. We estimated ordinary least squares (OLS)
regressions of any hospitalizations and any ER visits, and separately, the average number of
outpatient visits by dementia severity and quarter relative to diagnosis. Inpatient length of stays
and expenditures were estimated using two-part models due to the substantial share of zeros in
the data. The first part estimated the probability of having any inpatient length of stay or
expenditures during each quarter using a probit model, and the second part estimated the
magnitude of inpatient length of stay or expenditures conditional on having non-zero length of
stay or costs using a generalized linear model. Outpatient expenditures were estimated using
OLS regressions.

97
The models adjusted for age, age squared, sex, race/ethnicity (non-Hispanic white, nonHispanic black, Hispanic, and non-Hispanic other), education level (<12 years, 12-15 years, and
≥16 years), total wealth quartiles (relative to the position in this sample), and diagnosis with the
following comorbid conditions: diabetes, hypertension, hyperlipidemia, stroke, acute myocardial
infarction (AMI), and atrial fibrillation (ATF). Comorbid conditions were identified in TM
claims using diagnosis codes and the Chronic Conditions Data Warehouse algorithms in each
year. Standard errors were clustered at the individual level to account for within-person
correlation. Unadjusted quarterly health care utilization and costs by dementia severity at
diagnosis are provided in Supplementary Figure 4.2. Model estimates of any hospitalization, any
ER visit, number of outpatient visits and outpatient spending based on OLS regressions are
reported in Supplementary Table 4.4. Model estimates of inpatient length of stay and inpatient
spending based on two-part models are reported in Supplementary Table 4.5. All analyses were
conducted in STATA 15 (StataCorp LP, College Station, TX).
Sensitivity Analysis
Three sensitivity analyses were conducted to evaluate the robustness of results. First, we
estimated health care utilization and spending adjusting additionally for calendar year of
diagnosis and whether the HRS interview was before diagnosis in the models. Second, we
excluded individuals living in nursing homes from the analytical sample (based on self-report in
the closest HRS interview from the dementia diagnosis). Finally, we tested sensitivity of results
using an alternative time window which required the closest HRS interview to be up to 12
months before or after the dementia diagnosis.

98
RESULTS
Characteristics of Study Population by Dementia Severity at Diagnosis
Table 10 presents the characteristics of the sample by dementia severity at first diagnosis.
Among the 2015 respondents aged 70 and older with an incident dementia diagnosis, 376
(18.7%) were diagnosed at mild stage, 1338 (66.4%) at moderate stage, and 301(14.9%) at
severe stage of the disease. The mean (SD) age at diagnosis were 80.2 (5.9) years among persons
diagnosed at mild dementia, 83.9 (6.9) at moderate dementia, and 86.1 (6.9) at severe dementia.
Females were 57.7%, 64.1% and 68.8% of individuals diagnosed at mild, moderate and severe
stage of disease respectively. About 89% of persons diagnosed at mild stage are non-Hispanic
White persons with 10% non-Hispanic Black or Hispanic persons. In contrast, about 21% and
26% of persons diagnosed at moderate or severe stage respectively were non-Hispanic Black or
Hispanic persons.
Individuals diagnosed at mild stage were the most educated, with 28.7% of them having a
bachelor’s degree or higher. Only 12.5% and 9.6% of those diagnosed at moderate and severe
stage of disease completed college. Median wealth (2016 dollars) was $288,731 among persons
with mild disease at diagnosis. This compares to $114,410 and $56,852 among those with
moderate and severe disease at diagnosis, respectively. Persons diagnosed at moderate and
severe stage of disease were more likely to be without a spouse or partner (59.8% and 73.1%)
than persons diagnosed at mild stage disease (46.5%). About 43.5% of individuals diagnosed at
severe stage dementia were living in nursing homes. Respondents diagnosed at severe stage had
the highest rates of AMI (12.3%) and stroke (37.5%) but lower rates of diabetes (35.9%),
hypertension (86.4%), and hyperlipidemia (56.5%) compared to persons diagnosed at mild and
moderate stages of disease. The majority of persons diagnosed at mild stage disease reported no

99
difficulty with ADLs or IADLs (83.0% and 93.1%) while 58.5% and 93.0% of persons
diagnosed at severe stage reported at least three ADL difficulties and IADL difficulties. About
35.6% of persons diagnosed at severe stage died within four quarters following the dementia
diagnosis, compared to 24.2% and 27.4% among persons diagnosed at mild and moderate stage
disease, respectively.
Regression Models of Health Care Utilization and Costs
Outpatient Use and Spending
Figure 7 shows regression adjusted model predictions for the average number of
outpatient visits and outpatient spending for each category of dementia severity with other
covariates set at overall sample means. There were increases in outpatient visit counts and
spending in the quarter preceding diagnosis, and in the quarter of diagnosis. Outpatient visit
counts and spending declined in the quarters after diagnosis but remained elevated relative to
before diagnosis.
Outpatient Visits. Persons diagnosed at mild stage had higher levels of outpatient visits
relative to those diagnosed at moderate or severe stage disease. Four quarters prior to diagnosis,
persons diagnosed at mild stage had 6.3 visits. This level was fairly steady until the quarter prior
to diagnosis where number of visits was 8.1, a 29 percent increase. In the quarter of diagnosis,
persons diagnosed at mild stage had 13.1 visits, a further 62 percent increase from the quarter
before. Visits reduced to 8.6 the quarter after diagnosis, declined further but remained above prediagnosis levels four quarters after diagnosis (7.1 visits). Among persons diagnosed at moderate
stage outpatient visits increased from 6.8 to 12.2 visits (81% increase) in the quarter of diagnosis
and from 6.2 to 11.2 visits (81%) among persons diagnosed at severe stage disease.

100
In the four quarters prior to diagnosis, persons diagnosed at moderate stage dementia had
0.7 Q-4 (p<0.1), 0.9 Q-3 (p<0.05), 0.8 Q-2 (p<0.05) and 1.3 Q-1 (p<0.01) fewer outpatient visits
relative to persons diagnosed at mild stage. Persons diagnosed at severe stage dementia had 1.2 Q4 (p<0.01), 1.1 Q-3 (p<0.05), 1.3 Q-2 (p<0.01) and 1.9 Q-1 (p<0.001) fewer visits. In the quarter of
diagnosis, persons diagnosed at moderate and severe stage had 0.9 (p<0.1) and 1.9 (p<0.01)
fewer visits relative to mild stage. In the four quarters following diagnosis, while counts of
outpatient visits remained lower among persons diagnosed at later stages compared to mild stage,
the differentials across categories of dementia severity narrowed over time and were no longer
statistically significant thereafter (Figure 7a).
Outpatient Spending. Four quarters prior to diagnosis, persons diagnosed at mild stage
dementia had $1739 of costs associated with outpatient visits which were relatively stable until
the quarter before diagnosis ($2660). In the quarter of diagnosis, outpatient costs increased
further to $4498 among persons diagnosed at mild stage, a 69% increase from the quarter before
(Figure 7b). In the four quarters prior to diagnosis, outpatient spending was $183 Q-4 (p=0.34),
$421 Q-3 (p<0.1), $258 Q-2 (p=0.21) and $632 Q-1 (p<0.05) lower among persons diagnosed at
moderate stage disease relative to mild stage. Outpatient spending was $500 Q-4 (p<0.05), $578 Q3 (p<0.05), $419 Q-2 (p<0.1), and $923 Q-1 (p<0.01) lower among persons diagnosed at severe
stage disease relative to mild stage.
During the quarter of diagnosis, outpatient spending almost doubled from the quarter
before among persons diagnosed at later stages disease. Spending increased from $2028 to $3979
(96% increase) among persons diagnosed at moderate stage and from $1737 to $3454 (99%
increase) among those diagnosed at severe stage disease. In the quarter of diagnosis, outpatient
spending was $518 (p<0.1) and $1044 (p<0.01) lower among persons diagnosed at moderate and

101
severe stage relative to mild stage. Outpatient spending declined sharply after diagnosis for all
persons but remained lower for persons diagnosed at later stages of disease several quarters after
diagnosis. In the quarter immediately following the diagnosis, outpatient spending was $442
(p<0.1) and $578 (p<0.05) lower among persons diagnosed at moderate and severe stage relative
to mild stage. Among the most severe group, outpatient spending remained $502 Q2 (p<0.05),
$210 Q3 (p=0.44), and $405 Q4 (p<0.1) lower compared to mild group two to four quarters after
diagnosis.
Inpatient Use and Spending
Figure 8 shows regression adjusted model predictions for hospitalization rates, length of
stay, and spending for each category of dementia severity with other covariates set at overall
sample means. Across all levels of dementia severity at diagnosis, hospitalization rates, inpatient
length of stay, and inpatient spending increased in the quarter preceding the incident dementia
diagnosis and peaked in the quarter of diagnosis. Inpatient care utilization and expenditures
declined sharply after diagnosis but remained elevated for several quarters compared to prediagnosis period.
Hospitalization Rates. Among persons diagnosed at mild stage disease, hospitalization
rates increased from 18.3% in the quarter prior to diagnosis to 47.3% in the quarter of diagnosis.
Hospitalization rates increased from 15.9% to 50.7% among persons diagnosed at moderate stage
and from 16.4% to 54.0% among persons diagnosed at severe stage (Figure 8a). Hospitalization
rates were statistically indistinguishable across dementia stages in the quarters before (with
exception of Q-3 between mild and severe stage). During the quarter of diagnosis, hospitalization
rates were 3.4 (p=0.23) and 6.7 (p<0.1) percentage points higher among persons diagnosed at

102
moderate and severe stage disease relative to mild stage. Hospitalization rates were not
statistically different across categories of dementia severity in the quarters following diagnosis.
Inpatient Length of Stay. Among persons diagnosed at mild stage disease, inpatient
length of stay increased from 2.0 days in the quarter prior to 5.4 days (177% increase) in the
quarter of diagnosis. This compares to an increase from 1.5 to 5.6 days (260%) among persons
diagnosed at moderate stage and from 1.4 to 5.6 days (291%) among persons diagnosed at severe
stage (Figure 8b). In the four quarters before, the quarter of, and four quarters after diagnosis,
differences in average inpatient length of stay across dementia stages were not statistically
significant.
Inpatient Spending. From the quarter before to quarter of diagnosis, inpatient spending
rose from $3192 to $8491 (166% increase) among persons diagnosed at mild stage disease.
Inpatient spending increased from $2534 to $8831(248%) and from $2645 to $8169 (209%)
among persons diagnosed at moderate and severe stage disease, respectively (Figure 8c).
Average inpatient spending was statistically indistinguishable across categories of dementia
severity in the four quarters before (with exception of Q-3 between mild and severe stage), the
quarter of, and four quarters after diagnosis.
Emergency Room Use
Regression adjusted model predictions for ER rates were reported in Figure 9. Regardless
of dementia severity at diagnosis, quarterly ER rates were relatively stable four to two quarters
preceding the dementia diagnosis, ranging from 13% and 18%, and increased to between 22%
and 25% in the quarter prior to diagnosis. In the four quarters prior to diagnosis, ER rates were
statistically indistinguishable among persons diagnosed at different stages of dementia. During

103
the quarter of diagnosis, ER rates increased from 24.7% to 52.3% among persons diagnosed at
mild stage dementia. This compares to an increase from 24.0% to 54.4% among persons
diagnosed at moderate stage and from 22.3% to 57.2% among persons diagnosed at severe stage
disease. The differences across dementia stages in the quarter of diagnosis were not statistically
different. In the quarter immediately following diagnosis, ER rates declined to almost prediagnosis levels (21.4%) among persons diagnosed at mild stage disease. ER rates, however
remained elevated among those diagnosed at moderate and severe stage (28.2% and 30.6%). In
particular, compared to persons diagnosed at mild stage, ER rates were 6.7 (p<0.01) and 9.2
(p<0.01) percentage points higher among persons diagnosed at moderate and severe stage in the
quarter following diagnosis. ER rates remained 5.2 and 2.9 percentage points higher for the most
severe group relative to the mild group two to three quarters after diagnosis but the differences
were no longer statistically significant.
DISCUSSION
This study used HRS data linked with administrative TM claims data to quantify how
heterogeneity in dementia severity around the time of diagnosis is related to levels and trends in
health care utilization and costs over two years beginning with the year before dementia
diagnosis to the year after a dementia diagnosis. We found use of health care services and
spending increased a quarter before the diagnosis and increased most significantly in the quarter
of diagnosis. Both use and spending declined the quarter after diagnosis but remained elevated
relative to the second and earlier quarters before diagnosis. The large increases in use of health
care services and spending around the time of diagnosis we reported is consistent with findings
from prior studies (Coe et al., 2023; Hoffman et al., 2022; Jacobson et al., 2023; Lin et al., 2016;

104
White et al., 2019; C. W. Zhu et al., 2015). Although this general pattern was consistent for
persons diagnosed at mild, moderate and severe stages of dementia, there were both similarities
and differences in the levels and changes in use and costs for persons with different levels of
dementia severity around the time of diagnosis.
Acute care (i.e. inpatient and ER) use and costs as well as physician/outpatient use and
costs four quarters before a diagnosis, and most evident the quarter before diagnosis, is higher
for persons with mild dementia at time of diagnosis compared to persons with moderate or
severe dementia. While the higher level of use and costs associated with mild dementia relative
to moderate or severe holds for the quarter of diagnosis for physician/outpatient use, it reverses
for acute care use; persons with severe and moderate dementia at diagnosis have higher use of
acute care compared to persons with mild dementia during the quarter of diagnosis. The
likelihood of being hospitalized in the quarter of diagnosis is 3 and 7 percentage points higher
among persons diagnosed at moderate and severe stage disease compared to mild stage. Persons
diagnosed at moderate and severe stages also had higher ER rates relative to mild dementia –
about 2 and 5 percentage points higher in the quarter of diagnosis and 7 and 9 percentage points
higher in the quarter following a dementia diagnosis. Inpatient spending, however, is similar for
persons diagnosed at different dementia stages in the quarter of diagnosis and thereafter.
To the extent that some acute care services after a dementia diagnosis are considered lowvalue and potentially preventable (Bynum et al., 2004; Phelan et al., 2012), prior study has
pointed to potential for reducing dementia attributable acute care utilization via early detection
and diagnosis of the disease. The use and costs patterns of this study suggest some but limited
opportunity for reducing acute care service use of persons with moderate or severe dementia
around the time of diagnosis with little to no impact on health care spending. The results also

105
suggest that shifts to earlier diagnosis would increase use of physician services and spending on
outpatient care. In terms of spending, persons diagnosed at more advanced stage of dementia
may receive fewer procedures during hospitalizations compared to persons diagnosed at mild
stage disease. We find some supportive evidence for this explanation by examining procedures
received during hospitalizations for congestive heart failure, one of the most common reasons of
hospitalizations among persons with dementia (Phelan et al., 2012). We find that persons
diagnosed at the most advanced stage of dementia received fewer number and less invasive
procedures during their inpatient stays (Supplementary Table 4.6).
Persons who are diagnosed at mild stage of disease are more likely to be male and nonHispanic white persons, more likely to be college educated and have higher wealth than persons
diagnosed at moderate or severe stages. Despite rich controls for differences across the groups
defined by dementia severity around time of diagnosis in demographic characteristics, socioeconomic factors and health conditions that are associated with health care use and costs, there
may still remain unobserved differences. For example, persons diagnosed at mild stage dementia
may have higher preferences for health care use or better access to care and this high level of
interaction with the health care system provided clinical opportunities for early detection and
diagnosis of dementia. Among persons with less access or lower preference for health care use,
occurrence of an adverse health event combined with severe enough dementia symptoms may
have prompted a dementia diagnosis (Hoffman et al., 2022).
Prior study examined the relationship between dementia severity and health care costs
although not severity at diagnosis. A few studies reported health care costs increased as the
disease progresses (Leicht et al., 2011; Mauskopf et al., 2010; Zhu et al., 2008) while other
studies reported contradictory results (Ku et al., 2016; Michalowsky et al., 2018; Schwarzkopf et

106
al., 2011). For example, one study of persons in Taiwan, reported that persons with mild
dementia had higher drug expenditures compared to persons with more severe dementia (Ku et
al., 2016). Another study based on a community sample reported slightly higher inpatient,
outpatient, and rehabilitative costs for patients with mild stage dementia compared to later stages
(Schwarzkopf et al., 2011). There is a lack of generalizability of prior studies to older U.S. adults
due to non-representative or non-U.S. based samples, data limitations including small sample
sizes, cross-sectional data, and lack of comprehensive measures to capture all dimensions of
disease severity (cognitive, functional, behavioral) (Mauskopf et al., 2010; Rapp et al., 2012).
Moreover, the extant literature has focused exclusively on estimating medical costs based on
contemporary disease severity among prevalent dementia cases. These estimates are useful for
assessing use of health care resources over the course of the disease but have limited use for
quantifying the impact of timely dementia diagnosis. By considering dementia severity at
diagnosis, our findings enhance understanding of health care use and costs associated with
timing of dementia diagnosis.
This study has limitations. First, there may be misclassification in diagnoses of dementia
in claims data. Misclassification is reduced by use of a validated algorithm that includes two
claims of dementia diagnosis and dementia symptoms over time to account for rule-out
diagnoses. Second, our dementia severity measure is an approximation of the actual CDR. We
tested multiple models and assessed overall prediction accuracy (Supplementary materials
provide more details on model validation). Third, and as discussed above, there may be
unobserved differences associated with severity at diagnosis and health care use which may lead
to over-estimates for persons with mild dementia at diagnosis and under-estimates of persons
with severe. Fourth, our sample was drawn from beneficiaries enrolled in traditional Medicare.

107
The different benefit design, care organization and financial incentives to diagnosis health
conditions in Medicare Advantage may elicit different patterns of dementia diagnosis and health
care utilization and costs (Haye et al., 2023; Jacobson et al., 2023). Fifth, our measure of
dementia severity is based on the closest HRS interview to diagnosis date and may be a noisy
measure of severity at diagnosis. We reduce measurement error by adopting a time window
which requires the closest HRS interview to be up to 12 months before or up to 6 months after
the incident dementia. Furthermore to assess robustness of results, we adjusted the model for
calendar year and indicator for timing of severity measure before or after diagnosis and results
were robust (Supplementary Figure 4.3). We also tested robustness of results by excluding
individuals living in nursing homes (Supplementary Figure 4.4), and using an alternative time
window requirement (Supplementary Figure 4.5) and found consistent results. Finally, we only
examined use and costs of inpatient, ER, and outpatient care. Other components of medical care
and informal care were not examined in the current study.
Despite the limitations, the study offers new insight on health care use and cost
implications of early detection of dementia in the older U.S. population. Despite little evidence
to support that early detection and diagnosis of dementia will substantially reduce health care use
and spending around the time of diagnosis, improving detection for early diagnosis will likely
still have significant value to persons living with dementia and their families by providing
opportunities for better quality of life.

108
TABLES & FIGURES
Table 9. Sample Selection Criteria for the Study Population
Selection criterion N
HRS respondents with linked Traditional Medicare (TM) claims data
in 2000-2016
23,856
Three-year continuous TM enrollment and incident dementia 4,435
Aged 70 years and older 4,074
HRS interview up to 12 months before or up to 6 months after
incident dementia
2,281
HRS interview was no earlier than 2000 2,021
Complete information on cognitive or functional limitations 2,015
Notes: HRS, Health and Retirement Study; TM, traditional Medicare.

109
Table 10. Characteristics of Study Population by Dementia Severity at Diagnosis
Total
Sample
Mild
Dementia
(CDR<0.5)
Moderate
Dementia
(0.5≤CDR<2.5)
Severe
Dementia
(CDR≥2.5)
Sample, No. 2015 376 1338 301
Sociodemographic Characteristics
Mean Age in years at Dx (SD) 83.54 (6.94) 80.24 (5.93) 83.89 (6.89) 86.08 (6.86)
Female, % 63.57 57.71 64.05 68.77
Race/Ethnicity, %
Non-Hispanic White 78.26 88.56 76.98 71.10
Non-Hispanic Black 14.14 6.91 15.32 17.94
Hispanic 5.91 * 6.05 8.64
Education, %
Less than High School 34.14 13.30 36.47 49.83
High School/Some College 50.77 57.98 51.05 40.53
College and above 15.09 28.72 12.48 9.63
Median Total Wealth (2016$) 129,769 288,731 114,410 56,852
Single, % 59.31 46.54 59.79 73.09
Live in Nursing Home, % 13.35 * 9.72 43.52
Health and Functional Status
Comorbid Conditions, %
AMI 10.57 9.57 10.46 12.29
ATF 26.55 23.67 27.80 24.58
Stroke 32.80 26.33 33.56 37.54
Diabetes 39.06 37.23 40.28 35.88
Hypertension 89.63 87.50 90.96 86.38
Hyperlipidemia 70.97 78.72 72.05 56.48
ADL, %
No Difficulty 52.16 82.98 52.84 10.63
1-2 Difficulties 26.15 13.83 28.55 30.90
3-5 Difficulties 21.69 * 18.61 58.47
IADL, %
No Difficulty 48.14 93.09 46.34 *
1-2 Difficulties 26.55 6.91 36.47 *
3-5 Difficulties 25.31 * 17.19 93.02
Died in 4 Quarters after Dx, % 28.04 24.20 27.43 35.55
Survey before Diagnosis, % 70.17 78.99 71.45 53.49
Notes: Sample is HRS respondents with linked Traditional Medicare claims data in 2000-2016 who had an incident
dementia diagnosis in claims data that was verified over time by second claim, aged 70 years and older, and had an
HRS interview up to 12 months before or up to 6 months after incident dementia. AMI=acute myocardial infarction.
ATF=atrial fibrillation. ADL=activities of daily living. IADL=instrumental activities of daily living. Values with *
are censored (n<25).

110
Figure 7. Predicted Outpatient Visits and Costs Before and After Dementia Diagnosis by
Dementia Severity at Diagnosis
Notes: Predicted number of outpatient visits (Fig. 1a) and outpatient spending (Fig 1b) by dementia severity at
diagnosis and quarter relative to the date of incident dementia diagnosis, with other covariates at overall sample
means. OLS models were estimated, adjusting for age, age squared, sex, race, education, total wealth quartiles, and
comorbid conditions. Outpatient spending is converted to 2016 dollars. Based on Traditional Medicare claims and
HRS data.
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 6.25 6.48 6.59 8.06 13.12 8.60 7.69 7.33 7.12
Moderate (0.5-2.4) 5.59 5.55 5.76 6.75 12.19 7.99 7.27 6.92 6.92
Severe (2.5-4.4) 5.09 5.39 5.33 6.18 11.20 8.06 7.13 7.25 6.53
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
Outpatient Visits (#)
7a. Average Number of Outpatient Visits by Dementia Severity by
Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 1738.79 1963.51 1889.57 2659.98 4497.52 2718.70 2208.31 2002.57 1874.86
Moderate (0.5-2.4) 1555.87 1542.75 1631.15 2027.63 3979.34 2276.56 2081.47 1879.43 1768.27
Severe (2.5-4.4) 1238.39 1385.24 1470.94 1737.30 3454.01 2140.28 1706.81 1792.72 1469.84
0.00
500.00
1000.00
1500.00
2000.00
2500.00
3000.00
3500.00
4000.00
4500.00
5000.00
Outpatient Costs (2016$)
7b. Outpatient Costs by Dementia Severity by Quarter Relative to
Dementia Diagnosis

111
Figure 8. Predicted Inpatient Care Utilization and Costs Before and After Dementia
Diagnosis by Dementia Severity at Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 9.30 13.03 11.16 18.35 47.34 19.46 18.70 14.92 14.68
Moderate (0.5-2.4) 8.59 10.45 9.86 15.91 50.74 19.82 17.44 15.83 16.27
Severe (2.5-4.4) 10.13 7.47 8.14 16.44 53.99 21.87 16.08 16.66 12.20
0.00
10.00
20.00
30.00
40.00
50.00
60.00
Inpatient Stays (%)
8a. Percent of Individuals with an Inpatient Stay by Dementia Severity
by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 0.62 0.95 0.55 1.95 5.40 1.48 1.10 0.98 0.86
Moderate (0.5-2.4) 0.52 0.66 0.60 1.54 5.55 1.61 1.37 1.48 1.31
Severe (2.5-4.4) 0.54 0.56 0.82 1.44 5.63 1.77 1.21 1.19 1.22
0.00
1.00
2.00
3.00
4.00
5.00
6.00
Inpatient Length of Stay (Days)
8b. Inpatient Length of Stay by Dementia Severity by Quarter Relative
to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 1123.17 2208.38 1157.09 3191.83 8491.40 2847.65 2280.26 2094.84 2096.48
Moderate (0.5-2.4) 980.82 1285.56 1282.32 2534.43 8830.57 2600.10 2321.39 2579.22 2056.04
Severe (2.5-4.4) 892.51 887.20 997.02 2645.31 8168.50 2762.62 2579.45 2044.50 1636.50
0.00
1000.00
2000.00
3000.00
4000.00
5000.00
6000.00
7000.00
8000.00
9000.00
10000.00
Inpatient Costs (2016$)
8c. Inpatient Costs by Dementia Severity by Quarter Relative to
Dementia Diagnosis

112
Notes: Predicted hospitalization rates (Fig. 2a), inpatient length of stay (Fig. 2b), and inpatient spending (Fig. 2c) by
dementia severity at diagnosis and quarter relative to the date of incident dementia diagnosis, with other covariates
at overall sample means. OLS model was estimated for hospitalization rates and two-part models for inpatient length
of stay and spending, adjusting for age, age squared, sex, race, education, total wealth quartiles, and comorbid
conditions. Inpatient spending is converted to 2016 dollars. Based on Traditional Medicare claims and HRS data.

113
Figure 9. Predicted Emergency Room Utilization Before and After Dementia Diagnosis by
Dementia Severity at Diagnosis
Notes: Predicted ER rates by dementia severity at diagnosis and quarter relative to the date of incident dementia
diagnosis, with other covariates at overall sample means. OLS model was estimated, adjusting for age, age squared,
sex, race, education, total wealth quartiles, and comorbid conditions. Based on Traditional Medicare claims and
HRS data.
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 12.71 18.03 16.70 24.68 52.34 21.43 24.16 21.60 19.29
Moderate (0.5-2.4) 14.81 16.53 16.46 24.01 54.35 28.15 24.07 23.41 23.97
Severe (2.5-4.4) 16.36 18.02 15.03 22.34 57.22 30.61 29.37 24.48 17.55
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
ER Visits (%)
Percent of Individuals with any ER Visit by Dementia Severity by
Quarter Relative to Dementia Diagnosis

114
APPENDIX
Supplementary Table 4.1. ICD-9 and ICD-10 Codes Used to Define Dementia and
Dementia Symptoms
Code
Version
Code Description
ICD-9 290.0 Senile dementia, uncomplicated
290.10 Presenile dementia, uncomplicated
290.11 Presenile dementia w/ delirium
290.12 Presenile dementia w/ delusional features
290.13 Presenile dementia w/ depressive features
290.20 Senile dementia w/ delusional features
290.21 Senile dementia w/ depressive features
290.3 Senile dementia w/ delirium
290.40 Vascular dementia, uncomplicated
290.41 Vascular dementia, w/ delirium
290.42 Vascular dementia, w/ delusions
290.43 Vascular dementia, w/ depressed mood
294.0 Amnestic disorder in conditions cls-elsew
294.10 Dementia in conditions cls-elsew w/out behav disturbnc
294.11 Dementia in conditions cls-elsew w/ behav disturbnc
294.20 Dementia, unspecified, w/out behav disturbnc
294.21 Dementia, unspecified, w/ behav disturbnc
294.8 Oth persistent mental disorders due to conds cls-elsew
331.0 Alzheimer's Disease
331.11 Picks disease
331.19 Other frontotemporal dementia
331.2 Senile degeneration of brain
331.7 Cerebral degeneration in disease cls-elsew
797 Senility w/out mention of psychosis
780.93 Amnesia
784.3 Aphasia
784.69 Other symbolic dysfunctions, apraxia, agnosia
331.83 Mild cognitive impairment, so stated
ICD-10 F01.50 Vascular dementia, w/out behav disturbnc
F01.51 Vascular dementia, w/ behav disturbnc
F02.80 Dementia in conditions cls-elsew w/out behav disturbnc
F02.81 Dementia in conditions cls-elsew, w/ behav disturbnc
F03.90 Dementia, unspecified, w/out behav disturbnc
F03.91 Dementia, unspecified, w/ behav disturbnc
G13.8 Systemic atrophy aff cnsl in oth diseases classd elsew
G30.0 Alzheimer's disease with early onset
G30.1 Alzheimer's disease with late onset
G30.8 Other Alzheimer's disease
G30.9 Alzheimer's disease, unspecified
G31.01 Picks disease
G31.09 Other frontotemporal dementia
G31.1 Senile degeneration
G94 Oth disorders of brain in diseases classd elsew

115
Code
Version
Code Description
R41.81 Age-related cognitive decline
R54 Age-related physical debility
R41.1 Anterograde amnesia
R41.2 Retrograde amnesia
R41.3 Other amnesia
R47.01 Aphasia
R48.1 Agnosia
R48.2 Apraxia
R48.8 Other symbolic dysfunctions
G318.4 Mild cognitive impairment, so stated
Note: Codes are from inpatient, outpatient, home health care, skilled nursing facility and carrier claims.

116
Supplementary Table 4.2. Poisson Model Estimates of CDR in ADAMS, N=852
Parameter Estimate SE 95% CI p-value
Intercept -5.2210 3.6755 -12.4249 1.9829 0.1555
Age 0.0911 0.0871 -0.0796 0.2618 0.2956
Age squared -0.0005 0.0005 -0.0015 0.0005 0.3441
Female 0.0375 0.0754 -0.1103 0.1853 0.6192
Education
High school/some college 0.0636 0.0803 -0.0938 0.2209 0.4286
college and above 0.0929 0.1343 -0.1702 0.3560 0.4890
Race/ethnicity
Non-Hispanic Black 0.0253 0.0923 -0.1556 0.2062 0.7840
Hispanic -0.0021 0.1245 -0.2461 0.2419 0.9867
Non-Hispanic Other 0.0801 0.2007 -0.3134 0.4735 0.6899
Interview proxy status
Proxy 0.1913 0.0938 0.0074 0.3752 0.0415
Cognition status
Dementia 0.8206 0.1342 0.5576 1.0836 <.0001
CIND 0.5946 0.1232 0.3532 0.8359 <.0001
ADL
1-2 Difficulties 0.1201 0.1121 -0.0997 0.3399 0.2844
3-5 Difficulties 0.1981 0.1114 -0.0202 0.4165 0.0753
IADL
1-2 Difficulties 0.1568 0.1245 -0.0871 0.4008 0.2077
3-5 Difficulties 0.4563 0.1252 0.2110 0.7016 0.0003
Depression Status
Have depression -0.0957 0.1127 -0.3166 0.1252 0.3957
Missing 0.2848 0.1104 0.0684 0.5011 0.0099
Notes. Four ADAMS Wave A respondents with missing data on cognitive and functional limitations were excluded
(N=852). CIND=cognitively impairment but no dementia. IADL= instrumental activities of daily living.
ADL=activities of daily living.

117
Supplementary Figure 4.1. Observed and Predicted CDR Score in the ADAMS Sample
Notes: Sample is 852 respondents in the ADAMS. CDR, Clinical Dementia Rating.

118
Supplementary Table 4.3. Cross Tabulation of Observed and Predicted CDR in ADAMS,
N=852
Predicted CDR Score
Observed CDR Score
Poisson Model Including 0 0 1 2 3 4 Total
0 (CDR= 0) 151 (66) 74 (32) 3 (1) 0 0 228
1 (CDR= 0.5) 83 (23) 220 (60) 56 (15) 6 (2) 0 365
2 (CDR= 1) 2 (2) 48 (44) 51 (47) 7 (6) 0 108
3 (CDR= 2, 3) 0 4 (3) 48 (39) 59 (48) 13 (11) 124
4 (CDR= 4, 5) 0 0 4 (15) 15 (56) 8 (30) 27
Total 236 346 162 87 21 852
Overall accuracy, N (%) 489 (57%)
Simple Kappa 0.4017
Weighted Kappa 0.6062
P-B Adjusted Kappa 0.4674
Poisson Model Excluding 0 1 2 3 4 Total
1 (CDR= 0.5) 294 (81) 65 (18) 6 (2) 0 365
2 (CDR= 1) 45 (42) 61 (56) 2 (2) 0 108
3 (CDR= 2, 3) 3 (2) 49 (40) 62 (50) 10 (8) 124
4 (CDR= 4, 5) 0 4 (15) 15 (56) 8 (30) 27
Total 342 179 85 18
Overall accuracy, N (%) 425 (68%)
Simple Kappa 0.555
Weighted Kappa 0.6583
P-B Adjusted Kappa 0.7067
Note. Predicted CDR scores are rounded to the closest 0.25 place and binned.

119
Supplementary Figure 4.2. Unadjusted Health Care Utilization and Costs Before and After
Dementia Diagnosis by Dementia Severity at Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 7.45 11.17 9.31 16.49 45.48 17.44 16.62 12.62 12.16
Moderate (0.5-2.4) 9.12 10.99 10.39 16.44 51.27 20.19 17.50 15.77 16.05
Severe (2.5-4.4) 11.30 8.64 9.30 17.61 55.15 22.89 16.86 17.47 12.68
0.00
10.00
20.00
30.00
40.00
50.00
60.00
Inpatient Stays (%)
Unadjusted Percent of Individuals with an Inpatient Stay by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 0.64 0.93 0.55 1.94 5.25 1.48 1.15 0.98 0.82
Moderate (0.5-2.4) 0.64 0.79 0.72 1.75 5.66 1.75 1.46 1.56 1.37
Severe (2.5-4.4) 0.63 0.69 0.99 1.54 5.57 2.02 1.25 1.33 1.22
0.00
1.00
2.00
3.00
4.00
5.00
6.00
Inpatient LOS (Days)
Unadjusted Inpatient Length of Stay by Dementia Severity by Quarter
Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 1192.69 2192.62 1200.48 3255.41 8419.87 2924.49 2410.82 2170.69 2011.23
Moderate (0.5-2.4) 1177.27 1504.84 1515.21 2851.15 8919.68 2802.44 2443.53 2702.52 2148.07
Severe (2.5-4.4) 1035.66 1107.32 1162.76 2768.60 7861.48 3124.17 2693.63 2269.37 1550.67
0.00
1000.00
2000.00
3000.00
4000.00
5000.00
6000.00
7000.00
8000.00
9000.00
10000.00
Inpatienet Costs (2016$)
Unadjusted Inpatient Costs by Dementia Severity by Quarter Relative
to Dementia Diagnosis

120
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 6.39 6.62 6.73 8.20 13.26 8.74 7.81 7.43 7.16
Moderate (0.5-2.4) 5.67 5.63 5.84 6.83 12.27 8.04 7.27 6.90 6.87
Severe (2.5-4.4) 4.73 5.03 4.97 5.83 10.84 7.70 6.74 6.86 6.07
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
Outpatient Visits (#)
Unadjusted Average Number of Outpatient Visits by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 1875.10 2099.82 2025.88 2796.29 4633.83 2869.16 2354.50 2136.83 1993.37
Moderate (0.5-2.4) 1568.09 1554.98 1643.37 2039.86 3991.57 2282.35 2078.29 1869.30 1754.28
Severe (2.5-4.4) 1038.61 1185.46 1271.16 1537.52 3254.23 1948.99 1508.15 1600.26 1262.95
0.00
500.00
1000.00
1500.00
2000.00
2500.00
3000.00
3500.00
4000.00
4500.00
5000.00
Outpatient Costs (2016$)
Unadjusted Outpatient Costs by Dementia Severity by Quarter
Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild (<0.5) 10.64 15.96 14.63 22.61 50.27 19.19 21.85 19.09 16.55
Moderate (0.5-2.4) 15.40 17.12 17.04 24.59 54.93 28.57 24.18 23.37 23.78
Severe (2.5-4.4) 17.61 19.27 16.28 23.59 58.47 31.69 30.20 25.33 18.05
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
ER Visits (%)
Unadjusted Percent of Individuals with any ER Visit by Dementia
Severity by Quarter Relative to Dementia Diagnosis

121
Supplementary Table 4.4. OLS Model Estimates of Health Care Utilization and Spending
Any Inpatient
Stay
Any ER
Visit
Outpatient
Visits
Outpatient
Spending
(1) (2) (3) (4)
Lead 3_Mild 0.0372* 0.0532** 0.2314 224.7195
(0.0191) (0.0231) (0.3075) (187.1764)
Lead 2_Mild 0.0186 0.0399* 0.3431 150.7771
(0.0182) (0.0227) (0.3492) (194.9243)
Lead 1_Mild 0.0904*** 0.1197*** 1.8112*** 921.1901***
(0.0227) (0.0249) (0.4082) (245.1493)
Lag0_Mild 0.3803*** 0.3963*** 6.8723*** 2758.7266***
(0.0278) (0.0284) (0.4842) (274.0277)
Lag1_Mild 0.1016*** 0.0872*** 2.3519*** 979.9126***
(0.0230) (0.0241) (0.4384) (243.0076)
Lag2_Mild 0.0939*** 0.1145*** 1.4341*** 469.5216**
(0.0236) (0.0247) (0.4157) (225.5577)
Lag3_Mild 0.0562** 0.0889*** 1.0826*** 263.7785
(0.0221) (0.0277) (0.4110) (205.6145)
Lag4_Mild 0.0538** 0.0658** 0.8702** 136.0672
(0.0224) (0.0259) (0.4150) (199.4618)
Lead 4_Moderate -0.0072 0.0210 -0.6583* -182.9229
(0.0158) (0.0191) (0.3620) (190.1034)
Lead 3_Moderate 0.0115 0.0382** -0.7061* -196.0355
(0.0162) (0.0193) (0.3608) (190.6876)
Lead 2_Moderate 0.0055 0.0375* -0.4901 -107.6423
(0.0161) (0.0193) (0.3642) (195.4358)
Lead 1_Moderate 0.0661*** 0.1130*** 0.5024 288.8396
(0.0171) (0.0202) (0.3784) (201.8225)
Lag0_Moderate 0.4143*** 0.4164*** 5.9419*** 2240.5547***
(0.0192) (0.0212) (0.4120) (216.6184)
Lag1_Moderate 0.1052*** 0.1544*** 1.7379*** 537.7730***
(0.0177) (0.0206) (0.3815) (201.4488)
Lag2_Moderate 0.0814*** 0.1136*** 1.0155*** 342.6823*
(0.0176) (0.0205) (0.3799) (206.3889)
Lag3_Moderate 0.0653*** 0.1070*** 0.6672* 140.6390
(0.0176) (0.0208) (0.3886) (200.3884)
Lag4_Moderate 0.0697*** 0.1126*** 0.6663* 29.4829
(0.0177) (0.0210) (0.3899) (201.1480)
Lead 4_Severe 0.0083 0.0365 -1.1662*** -500.3985**
(0.0233) (0.0280) (0.4464) (210.7399)
Lead 3_Severe -0.0183 0.0531* -0.8606* -353.5448
(0.0217) (0.0286) (0.4751) (225.1615)
Lead 2_Severe -0.0117 0.0232 -0.9204** -267.8470
(0.0221) (0.0271) (0.4622) (229.0617)

122
Any Inpatient
Stay
Any ER
Visit
Outpatient
Visits
Outpatient
Spending
(1) (2) (3) (4)
Lead 1_Severe 0.0714*** 0.0963***
-0.0666
-1.4935
(0.0259) (0.0296) (0.4889) (251.4883)
Lag0_Severe 0.4468*** 0.4451*** 4.9467*** 1715.2190***
(0.0319) (0.0331) (0.5680) (265.6625)
Lag1_Severe 0.1257*** 0.1790*** 1.8089*** 401.4922
(0.0284) (0.0322) (0.5244) (245.1405)
Lag2_Severe 0.0677** 0.1666*** 0.8792
*
-31.9772
(0.0273) (0.0334) (0.5007) (222.3761)
Lag3_Severe 0.0736** 0.1177*** 0.9985
* 53.9325
(0.0286) (0.0331) (0.5231) (268.9424)
Lag4_Severe 0.0289 0.0484 0.2836
-268.9460
(0.0276) (0.0322) (0.5022) (237.6379)
Age
-0.0166
-0.0259**
-0.5914
*
-416.1997***
(0.0106) (0.0125) (0.3084) (152.6725)
Age2 0.0001 0.0002** 0.0031
* 2.1771**
(0.0001) (0.0001) (0.0018) (0.8884)
Female B
-0.0078 0.0026
-0.0778
-95.2625
(0.0076) (0.0092) (0.2149) (109.5392)
Non
-Hispanic Black 0.0105
-0.0003
-0.2287 279.7908
(0.0119) (0.0151) (0.3218) (199.9227)
Hispanic 0.0155 0.0220
-0.1242
-101.5713
(0.0159) (0.0219) (0.4161) (194.0522)
Non
-Hispanic Other
-0.0223 0.0042
-0.3077
-166.1032
(0.0282) (0.0384) (0.9967) (481.8806)
Education (12
-15 years)
-0.0184**
-0.0183
* 0.2702 152.0791
(0.0092) (0.0111) (0.2429) (124.4394)
Education (≥16 years)
-0.0162
-0.0220 0.4588 158.6312
(0.0120) (0.0144) (0.3279) (163.0928)
2nd Wealth Quartile
-0.0078
-0.0103
-0.1763
-103.0661
(0.0101) (0.0126) (0.2837) (142.2214)
3rd Wealth Quartile
-0.0257**
-0.0308**
-0.4940
*
-214.6550
(0.0106) (0.0130) (0.2790) (130.8666)
4th Wealth Quartile
-0.0433***
-0.0564***
-0.6757**
-261.3975
*
(0.0109) (0.0130) (0.3066) (142.6078)
Diabetes 0.0424*** 0.0440*** 1.5633*** 738.5414***
(0.0080) (0.0096) (0.2226) (112.9653)
Hypertension 0.0604*** 0.0704*** 1.4557*** 424.7218***
(0.0095) (0.0124) (0.2763) (120.5205)
Hyperlipidemia
-0.0229***
-0.0187
* 0.8964*** 335.5445***
(0.0083) (0.0100) (0.2097) (98.2168)
Stroke 0.0587*** 0.0588*** 1.2017*** 384.6056***
(0.0083) (0.0097) (0.2255) (110.7966)

123
Any Inpatient
Stay
Any ER
Visit
Outpatient
Visits
Outpatient
Spending
(1) (2) (3) (4)
AMI 0.0496*** 0.0508*** 0.2628 179.4635
(0.0132) (0.0169) (0.3514) (174.9302)
ATF 0.0674*** 0.0699*** 2.2937*** 484.0399***
(0.0088) (0.0104) (0.2472) (110.0841)
Observations 16842 16842 16842 16842
R2 0.1316 0.1011 0.1432 0.0883
Notes: The sample is restricted to HRS respondents with linked Traditional Medicare (TM) claims data in 2000-
2016 who had an incident dementia diagnosis in claims data that was verified over time by second diagnosis claim,
who was aged 70 years and older, and had an HRS interview up to 12 months before or up to 6 months after incident
dementia that occurred no earlier than 2000. Respondents deceased after dementia diagnosis were dropped from
subsequent quarters in the analysis. * p < 0.10, ** p < 0.05, *** p < 0.01

124
Supplementary Table 4.5. Two-Part Model Estimates of Inpatient Length of Stay and
Spending
Inpatient Length of Stay Inpatient Spending
(1) (2)
First Part: Probit
Lead 3_Mild 0.2437** 0.2435**
(0.1194) (0.1195)
Lead 2_Mild 0.1361 0.1363
(0.1229) (0.1228)
Lead 1_Mild 0.4836*** 0.4836***
(0.1231) (0.1231)
Lag0_Mild 1.3819*** 1.3820***
(0.1142) (0.1142)
Lag1_Mild 0.5270*** 0.5269***
(0.1203) (0.1204)
Lag2_Mild 0.5099*** 0.5097***
(0.1256) (0.1256)
Lag3_Mild 0.3452*** 0.3449***
(0.1290) (0.1290)
Lag4_Mild 0.3152** 0.3152**
(0.1345) (0.1345)
Lead 4_Moderate 0.0028 0.0100
(0.1101) (0.1101)
Lead 3_Moderate 0.1188 0.1252
(0.1094) (0.1093)
Lead 2_Moderate 0.0883 0.0906
(0.1096) (0.1096)
Lead 1_Moderate 0.3786*** 0.3906***
(0.1074) (0.1073)
Lag0_Moderate 1.4294*** 1.4359***
(0.1044) (0.1044)
Lag1_Moderate 0.5387*** 0.5413***
(0.1068) (0.1068)
Lag2_Moderate 0.4522*** 0.4546***
(0.1080) (0.1079)
Lag3_Moderate 0.3910*** 0.3934***
(0.1094) (0.1094)
Lag4_Moderate 0.3959*** 0.3977***
(0.1096) (0.1096)
Lead 4_Severe 0.1265 0.1303
(0.1398) (0.1398)
Lead 3_Severe -0.0284 -0.0248
(0.1455) (0.1455)
Lead 2_Severe 0.0063 0.0100

125
Inpatient Length of Stay Inpatient Spending
(1) (2)
(0.1436) (0.1436)
Lead 1_Severe 0.4120*** 0.4158***
(0.1318) (0.1318)
Lag0_Severe 1.5123*** 1.5164***
(0.1243) (0.1243)
Lag1_Severe 0.6148*** 0.6186***
(0.1304) (0.1304)
Lag2_Severe 0.4058*** 0.4097***
(0.1381) (0.1382)
Lag3_Severe 0.4253*** 0.4291***
(0.1409) (0.1409)
Lag4_Severe 0.2460 0.2499
(0.1536) (0.1537)
Age
-0.0598
-0.0665
(0.0445) (0.0444)
Age2 0.0004 0.0004
(0.0003) (0.0003)
Female B
-0.0401
-0.0401
(0.0323) (0.0323)
Non
-Hispanic Black 0.0367 0.0382
(0.0471) (0.0469)
Hispanic 0.0459 0.0433
(0.0606) (0.0606)
Non
-Hispanic Other
-0.0867
-0.0904
(0.1238) (0.1239)
Education (12
-15 years)
-0.0829**
-0.0819**
(0.0373) (0.0373)
Education (≥16 years)
-0.0704
-0.0645
(0.0515) (0.0513)
2nd Wealth Quartile
-0.0299
-0.0295
(0.0399) (0.0398)
3rd Wealth Quartile
-0.1097**
-0.1145**
(0.0448) (0.0448)
4th Wealth Quartile
-0.1915***
-0.1941***
(0.0482) (0.0482)
Diabetes 0.1752*** 0.1738***
(0.0327) (0.0327)
Hypertension 0.3587*** 0.3602***
(0.0567) (0.0567)
Hyperlipidemia
-0.0984***
-0.0976***
(0.0359) (0.0359)
Stroke 0.2357*** 0.2358***
(0.0324) (0.0324)

126
Inpatient Length of Stay Inpatient Spending
(1) (2)
AMI 0.1886*** 0.1869***
(0.0473) (0.0472)
ATF 0.2670*** 0.2703***
(0.0336) (0.0336)
Second Part: GLM
Lead 3_Mild 0.0381 4048.6580
(3.7169) (6186.3473)
Lead 2_Mild
-2.4281
-2884.7970
(3.1249) (4229.0695)
Lead 1_Mild 3.2596 3849.9817
(3.3901) (4922.4891)
Lag0_Mild 3.1891 2956.1211
(3.1133) (4247.1053)
Lag1_Mild 0.0463 834.6884
(3.1512) (4634.3434)
Lag2_Mild
-1.8627
-1944.5449
(3.0442) (4275.2792)
Lag3_Mild
-0.9890 371.3081
(3.1526) (5855.4521)
Lag4_Mild
-1.5391 1118.4761
(3.1953) (6461.2095)
Lead 4_Moderate
-1.2889
-2099.8501
(3.0125) (4076.5728)
Lead 3_Moderate
-1.0731
-1313.5239
(3.0573) (4106.2023)
Lead 2_Moderate
-1.3750
-494.4363
(3.0389) (4204.9363)
Lead 1_Moderate 2.4610 2265.6358
(3.0696) (4191.4713)
Lag0_Moderate 3.0672 2876.1271
(2.9912) (3994.2541)
Lag1_Moderate 0.6027
-806.1048
(3.0047) (4002.0874)
Lag2_Moderate 0.2923
-610.5824
(3.0190) (4041.9893)
Lag3_Moderate 1.8195 2490.8150
(3.0593) (4193.8003)
Lag4_Moderate 0.6185
-1085.6850
(3.0400) (4114.6584)
Lead 4_Severe
-2.3847
-5487.4012
(3.1118) (4187.2397)
Lead 3_Severe
-0.3490
-2530.7106
(3.5694) (4612.7599)

127
Inpatient Length of Stay Inpatient Spending
(1) (2)
Lead 2_Severe 2.5774 -1891.7645
(3.6361) (4253.4567)
Lead 1_Severe 1.2916 2331.1933
(3.2777) (5590.0053)
Lag0_Severe 2.5343 584.3773
(3.0403) (4091.3484)
Lag1_Severe 0.4676 -1457.8348
(3.2349) (4424.8067)
Lag2_Severe -0.1414 2064.9889
(3.1722) (5342.7263)
Lag3_Severe -0.5157 -1800.1886
(3.1397) (4216.9257)
Lag4_Severe 2.2974 -952.9448
(4.4079) (4682.9547)
Age 0.1998 -492.5873
(0.5975) (1055.7893)
Age2 -0.0021 1.1998
(0.0035) (6.2146)
Female
B
-0.0245 -441.1469
(0.4355) (819.7228)
Non-Hispanic Black 2.6917*** 5297.8778***
(0.7155) (1397.3194)
Hispanic 1.9482** 5985.5548***
(0.8231) (1718.9167)
Non-Hispanic Other 2.9931 4042.5025
(1.9877) (3266.7319)
Education (12-15 years) -0.2532 599.7889
(0.4581) (853.7355)
Education (≥16 years) -0.0527 115.9412
(0.6683) (1137.9150)
2nd Wealth Quartile 0.6679 -170.8204
(0.5362) (952.2456)
3rd Wealth Quartile 0.8196 1752.8918
(0.5565) (1092.4300)
4th Wealth Quartile 0.7521 1701.6155
(0.6144) (1097.3672)
Diabetes 0.7314* 705.2835
(0.4001) (755.8677)
Hypertension -1.3034 -3768.8430
(1.2373) (3138.1533)
Hyperlipidemia 0.3847 1881.1240**
(0.4620) (872.8096)
Stroke 0.7072* 1364.6624

128
Inpatient Length of Stay Inpatient Spending
(1) (2)
(0.4189) (849.3111)
AMI 0.2579 486.9245
(0.5775) (1104.8990)
ATF 0.7611* 2088.8723**
(0.4383) (861.6704)
Observations 16842 16842
Notes: The sample is restricted to HRS respondents with linked Traditional Medicare (TM) claims data in 2000-
2016 who had an incident dementia diagnosis in claims data that was verified over time by second diagnosis claim,
who was aged 70 years and older, and had an HRS interview up to 12 months before or up to 6 months after incident
dementia that occurred no earlier than 2000. Respondents deceased after dementia diagnosis were dropped from
subsequent quarters in the analysis. * p < 0.10, ** p < 0.05, *** p < 0.01

129
Supplementary Table 4.6. Procedures Received during Hospitalizations for Congestive
Heart Failure in the Quarter of Diagnosis by Dementia Severity at Diagnosis
Mild
Dementia
(CDR<0.5)
Moderate
Dementia
(0.5≤CDR<2.5)
Severe
Dementia
(CDR≥2.5)
Average #
Procedures 0.69 0.78 0.22
Name of
Procedures
• Diagnostic ultrasound of
heart
• Coronary Arteriography
Using Two Catheters
• Angiocardiography of left
heart structures
• Left Heart Cardiac
Catheterization
• Non-invasive mechanical
ventilation
• Thoracentesis
• Arteriography of other intraabdominal arteries
• Central Venous Catheter
Placement
• Diagnostic ultrasound of
heart
• Angiocardiography of left
heart structures
• Combined right and left
heart cardiac
catheterization
• Electrographic monitoring
• Pulmonary scan
• C.A.T. scan of thorax
• Arteriography of
pulmonary arteries
• Injection of anticoagulant
• Phlebography of other
intrathoracic veins using
contrast material
• Contrast Aortography
• Diagnostic Ultrasound of
Peripheral Vascular
System
• Angiocardiography of
Right Heart Structures
• Left Heart Cardiac
Catheterization
• Implantation of cardiac
resynchronization
pacemaker
• Other Endoscopy of Small
Intestine
• Transfusion of
packed cells
• Venous
Catheterization
Note: Hospitalizations for congestive heart failure are identified by the primary diagnosis on the claim.

130
-4 -3 -2 -1 0 1 2 3 4
Mild 8.92 12.64 10.78 17.96 46.95 19.11 18.36 14.60 14.38
Moderate 8.52 10.39 9.80 15.85 50.68 19.80 17.46 15.86 16.31
Severe 10.70 8.05 8.71 17.02 54.56 22.47 16.71 17.30 12.92
0.00
10.00
20.00
30.00
40.00
50.00
60.00
Inpaient Stays (%)
Percent of Individuals with an Inpatient Stay by Dementia
Severity and Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 0.59 0.92 0.51 1.84 5.23 1.42 1.04 0.96 0.84
Moderate 0.51 0.66 0.60 1.53 5.51 1.61 1.37 1.47 1.30
Severe 0.57 0.60 0.86 1.53 5.83 1.86 1.31 1.25 1.36
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
Inpatient LOS (Days)
Inpatient Length of Stay by Dementia Severity and
Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1093.0 2153.8 1118.0 3056.7 8342.6 2761.6 2208.8 2091.0 2052.9
Moderate 973.85 1286.1 1269.8 2492.3 8756.1 2586.7 2324.6 2603.0 2064.6
Severe 907.56 897.29 1042.5 2718.1 8285.9 2819.3 2696.5 2156.5 1766.7
0.00
2000.00
4000.00
6000.00
8000.00
10000.00
Inpatient Costs (2016$)
Inpatient Costs by Dementia Severity and Quarter
Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 12.59 17.91 16.58 24.56 52.22 21.34 24.08 21.54 19.25
Moderate 14.77 16.49 16.42 23.97 54.31 28.14 24.09 23.43 24.00
Severe 16.54 18.20 15.21 22.52 57.40 30.81 29.61 24.72 17.83
0.00
10.00
20.00
30.00
40.00
50.00
60.00
70.00
ER Visits (%)
Percent of Individuals with any ER Visit by Dementia
Severity and Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 6.24 6.47 6.59 8.05 13.12 8.60 7.69 7.34 7.13
Moderate 5.58 5.54 5.75 6.75 12.18 7.99 7.27 6.92 6.92
Severe 5.10 5.41 5.35 6.20 11.22 8.08 7.16 7.28 6.57
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
Outpatient Visits (#)
Average Number of Outpatient Visist by Dementia
Severity and Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1717.2 1941.9 1868.0 2638.4 4475.9 2704.9 2196.9 1993.6 1869.0
Moderate 1547.3 1534.2 1622.6 2019.1 3970.8 2274.7 2085.1 1885.3 1775.7
Severe 1271.7 1418.5 1504.2 1770.6 3487.3 2178.8 1751.6 1838.1 1524.3
0.00
500.00
1000.00
1500.00
2000.00
2500.00
3000.00
3500.00
4000.00
4500.00
5000.00
Outpatient Costs (2016$)
Outpatient Costs by Dementia Severity by Quarter
Relative to Dementia Diagnosis
Supplementary Figure 4.3. Predicted Health Care Utilization and Costs Before and After
Dementia Diagnosis by Dementia Severity at Diagnosis, Adjusting Additionally for
Calendar Year of Diagnosis and Whether HRS before Diagnosis
Notes: Predicted health care use and costs by dementia severity at diagnosis and quarter relative to the date of
incident dementia diagnosis. All models adjust for age, age squared, sex, race, education, total wealth quartiles,
comorbid conditions (diabetes, hypertension, hyperlipidemia, stroke, acute myocardial infarction, and atrial
fibrillation), calendar year of diagnosis, and whether HRS is before the diagnosis. Inpatient and outpatient costs are
converted to 2016$. Based on Traditional Medicare claims and HRS data.

131
-4 -3 -2 -1 0 1 2 3 4
Mild 9.46 13.00 11.36 17.61 47.23 18.86 18.88 15.07 14.83
Moderate 7.82 9.89 9.48 15.27 49.87 20.05 17.86 15.75 16.56
Severe 12.21 6.33 6.92 13.39 51.03 23.29 13.47 18.66 11.68
0.00
10.00
20.00
30.00
40.00
50.00
60.00
Inpaient Stays (%)
Percent of Individuals with an Inpatient Stay by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 0.62 0.96 0.57 1.91 5.45 1.46 1.12 0.99 0.88
Moderate 0.44 0.62 0.54 1.47 5.47 1.62 1.39 1.49 1.36
Severe 0.46 0.22 0.44 1.09 5.74 1.91 0.82 1.41 1.38
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
Inpatient LOS (Days)
Inpatient Length of Stay by Dementia Severity by Quarter
Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1131.4 2208.1 1180.5 3110.9 8559.0 2850.0 2306.6 2125.8 2134.5
Moderate 873.03 1250.8 1186.7 2361.9 8640.1 2630.6 2376.9 2584.0 2176.6
Severe 846.68 528.23 653.05 1901.4 8342.2 3106.4 1787.6 2251.3 1774.2
0.00
1000.00
2000.00
3000.00
4000.00
5000.00
6000.00
7000.00
8000.00
9000.00
10000.00
Inpatient Costs (2016$)
Inpatient Costs by Dementia Severity by Quarter Relative
to Dementia Diagnosis
Mild 12.52 17.96 16.60 24.21 52.47 21.58 24.22 21.30 19.63
Moderate 14.19 16.26 15.77 22.97 53.60 28.36 24.36 23.01 24.38
Severe 19.13 17.96 13.84 20.31 55.60 29.90 29.20 25.56 15.45
0.00
10.00
20.00
30.00
40.00
50.00
60.00
ER Visits (%)
Percent of Individuals with any ER Visit by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 6.17 6.41 6.54 7.92 13.07 8.56 7.61 7.25 7.00
Moderate 5.43 5.35 5.52 6.49 11.89 7.84 7.04 6.72 6.76
Severe 4.90 5.15 4.93 5.56 10.53 7.57 6.55 7.07 6.25
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
Outpatient Visits (#)
Average Number of Outpatient Visist by Dementia
Severity and Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1768.0 1987.1 1917.3 2682.7 4516.3 2755.3 2238.3 1987.3 1867.9
Moderate 1479.2 1488.0 1576.6 1975.5 3922.2 2269.3 2060.2 1880.0 1798.3
Severe 1258.2 1366.1 1356.9 1590.2 3206.8 2049.5 1651.3 1872.1 1483.6
0.00
500.00
1000.00
1500.00
2000.00
2500.00
3000.00
3500.00
4000.00
4500.00
5000.00
Outpatient Costs (2016$)
Outpatient Costs by Dementia Severity and Quarter
Relative to Dementia Diagnosis
Supplementary Figure 4.4. Predicted Health Care Utilization and Costs Before and After
Dementia Diagnosis by Dementia Severity at Diagnosis, Excluding Individuals Living in
Nursing Homes
Notes: Predicted health care use and costs by dementia severity at diagnosis and quarter relative to the date of
incident dementia diagnosis. The sample excluded persons living in nursing homes. All models adjust for age, age
squared, sex, race, education, total wealth quartiles, and comorbid conditions (diabetes, hypertension,
hyperlipidemia, stroke, acute myocardial infarction, and atrial fibrillation). Inpatient and outpatient costs are
converted to 2016$. Based on Traditional Medicare claims and HRS data.

132
-4 -3 -2 -1 0 1 2 3 4
Mild 8.68 12.70 11.04 17.42 45.56 18.06 17.05 15.10 13.82
Moderate 8.73 9.96 10.03 16.18 49.41 19.37 16.24 15.87 15.96
Severe 9.21 8.73 9.21 15.05 53.98 21.43 20.35 17.62 13.44
0.00
10.00
20.00
30.00
40.00
50.00
60.00
Inpaient Stays (%)
Percent of Individuals with an Inpatient Stay by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 0.69 0.98 0.63 2.10 5.03 1.48 1.14 1.17 0.89
Moderate 0.59 0.69 0.68 1.62 5.39 1.59 1.29 1.54 1.34
Severe 0.68 0.74 0.93 1.34 5.46 1.86 1.63 1.47 1.23
0.00
1.00
2.00
3.00
4.00
5.00
6.00
Inpatient LOS (Days)
Inpatient Length of Stay by Dementia Severity by Quarter
Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1197.5 2118.5 1236.4 3389.4 7958.6 2708.1 2248.3 2569.3 2012.0
Moderate 1082.2 1330.8 1393.3 2707.9 8535.6 2588.1 2182.5 2597.7 2153.5
Severe 1044.2 1224.6 1253.5 2357.3 7981.8 2914.6 3261.4 2546.9 1824.0
0.00
1000.00
2000.00
3000.00
4000.00
5000.00
6000.00
7000.00
8000.00
9000.00
Inpatient Costs (2016$)
Inpatient Costs by Dementia Severity by Quarter Relative
to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 12.13 17.57 16.38 22.77 50.90 20.27 22.86 22.00 18.69
Moderate 14.91 16.27 16.82 23.53 52.88 27.90 23.65 23.05 23.00
Severe 15.25 17.69 15.74 21.82 56.13 29.02 30.46 24.68 20.23
0.00
10.00
20.00
30.00
40.00
50.00
60.00
ER Visits (%)
Percent of Individuals with any ER Visit by Dementia
Severity by Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 6.13 6.47 6.51 7.91 12.98 8.40 7.46 7.32 6.93
Moderate 5.62 5.57 5.88 6.79 12.13 8.07 7.28 7.02 6.95
Severe 4.93 5.24 5.44 5.90 11.06 7.99 7.58 7.49 6.61
0.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
Outpatient Visits (#)
Average Number of Outpatient Visist by Dementia
Severity and Quarter Relative to Dementia Diagnosis
-4 -3 -2 -1 0 1 2 3 4
Mild 1682.1 1937.2 1877.7 2575.3 4306.8 2534.0 2286.5 2026.6 1826.6
Moderate 1524.9 1509.1 1661.0 2027.9 3951.3 2288.1 2061.2 1932.9 1789.5
Severe 1209.9 1296.2 1455.1 1660.6 3321.5 2047.1 1765.1 1816.5 1495.1
0.00
500.00
1000.00
1500.00
2000.00
2500.00
3000.00
3500.00
4000.00
4500.00
5000.00
Outpatient Costs (2016$)
Outpatient Costs by Dementia Severity and Quarter
Relative to Dementia Diagnosis
Supplementary Figure 4.5. Predicted Health Care Utilization and Costs Before and After
Dementia Diagnosis by Dementia Severity at Diagnosis, Using Alternative Time Window
Notes: Predicted health care use and costs by dementia severity at diagnosis and quarter relative to the date of
incident dementia diagnosis. The sample is restricted to HRS respondents with an HRS interview up to 12 months
before or up to 12 months after incident dementia that occurred no earlier than 2000. All models adjust for age, age
squared, sex, race, education, total wealth quartiles, and comorbid conditions (diabetes, hypertension,
hyperlipidemia, stroke, acute myocardial infarction, and atrial fibrillation). Inpatient and outpatient spending are
converted to 2016$. Based on Traditional Medicare claims and HRS data.

133
REFERENCES
Acevedo, A., Mullin, B. O., Progovac, A. M., Caputi, T. L., McWilliams, J. M., & Cook, B. L.
(2021). Impact of the Medicare Shared Savings Program on utilization of mental health
and substance use services by eligibility and race/ethnicity. Health Serv Res, 56(4), 581-
591. https://doi.org/10.1111/1475-6773.13625
Alzheimer’s Association. (2023). 2023 Alzheimer's disease facts and figures. Alzheimers
Dement, 19(4), 1598-1695. https://doi.org/10.1002/alz.13016
Amjad, H., Roth, D. L., Sheehan, O. C., Lyketsos, C. G., Wolff, J. L., & Samus, Q. M. (2018).
Underdiagnosis of dementia: an observational study of patterns in diagnosis and
awareness in US older adults. Journal of general internal medicine, 33, 1131-1138.
https://doi.org/10.1007/s11606-018-4377-y
Anderson, R. E., Ayanian, J. Z., Zaslavsky, A. M., & McWilliams, J. M. (2014). Quality of care
and racial disparities in medicare among potential ACOs. J Gen Intern Med, 29(9), 1296-
1304. https://doi.org/10.1007/s11606-014-2900-3
Anderson, T. S., Marcantonio, E. R., McCarthy, E. P., & Herzig, S. J. (2020). National Trends in
Potentially Preventable Hospitalizations of Older Adults with Dementia. Journal of the
American Geriatrics Society, 68(10), 2240-2248.
https://doi.org/https://doi.org/10.1111/jgs.16636
Baller, J. B., McGinty, E. E., Azrin, S. T., Juliano-Bult, D., & Daumit, G. L. (2015). Screening
for cardiovascular risk factors in adults with serious mental illness: a review of the
evidence. BMC psychiatry, 15(1), 1-13.
Barnett, J. H., Lewis, L., Blackwell, A. D., & Taylor, M. (2014). Early intervention in
Alzheimer’s disease: a health economic study of the effects of diagnostic timing. BMC
Neurology, 14(1), 101. https://doi.org/10.1186/1471-2377-14-101
Brown, M. T., & Wolf, D. A. (2018). Estimating the prevalence of serious mental illness and
dementia diagnoses among Medicare beneficiaries in the health and retirement study.
Research on aging, 40(7), 668-686.
Busch, A., Huskamp, H., & McWilliams, J. (2016). Early Efforts By Medicare Accountable Care
Organizations Have Limited Effect On Mental Illness Care And Management. Health
Affairs, 35, 1247-1256. https://doi.org/10.1377/hlthaff.2015.1669
Busch, A. B., Huskamp, H. A., Kreider, A. R., & McWilliams, J. M. (2017). Medicare
accountable care organizations and antidepressant use by patients with depression.
Psychiatric services, 68(11), 1193-1196.
Bynum, J. P. W., Rabins, P. V., Weller, W., Niefeld, M., Anderson, G. F., & Wu, A. W. (2004).
The Relationship Between a Dementia Diagnosis, Chronic Illness, Medicare
Expenditures, and Hospital Use [https://doi.org/10.1111/j.1532-5415.2004.52054.x].

134
Journal of the American Geriatrics Society, 52(2), 187-194.
https://doi.org/https://doi.org/10.1111/j.1532-5415.2004.52054.x
Cai, L., & Huang, J. (2018). Schizophrenia and risk of dementia: a meta-analysis study.
Neuropsychiatric disease and treatment, 14, 2047.
Callahan, C. M., Arling, G., Tu, W., Rosenman, M. B., Counsell, S. R., Stump, T. E., & Hendrie,
H. C. (2012). Transitions in care for older adults with and without dementia. J Am
Geriatr Soc, 60(5), 813-820. https://doi.org/10.1111/j.1532-5415.2012.03905.x
Chen, Y., Tysinger, B., Crimmins, E., & Zissimopoulos, J. M. (2019). Analysis of dementia in
the US population using Medicare claims: Insights from linked survey and administrative
claims data. Alzheimers Dement (N Y), 5, 197-207.
https://doi.org/10.1016/j.trci.2019.04.003
Chesney, E., Goodwin, G. M., & Fazel, S. (2014). Risks of all‐cause and suicide mortality in
mental disorders: a meta‐review. World psychiatry, 13(2), 153-160.
Chukmaitov, A. S., Harless, D. W., Bazzoli, G. J., & Deng, Y. (2019). Factors associated with
hospital participation in Centers for Medicare and Medicaid Services' Accountable Care
Organization programs. Health Care Manage Rev, 44(2), 104-114.
https://doi.org/10.1097/hmr.0000000000000182
CMS. (2014). Medicare Shared Savings Program: Shared Savings and Losses and Assignment
Methodology. Washington, DC: Centers for Medicare and Medicaid Services. Retrieved
Febuary 2024 from https://www.cms.gov/Medicare/Medicare-Fee-for-ServicePayment/sharedsavingsprogram/Downloads/Shared-Savings-Losses-AssignmentSpec.pdf
CMS. (2015). Fast Facts: All Medicare Shared Savings Program ACOs and Pioneer ACOs.
Washington, DC: Centers for Medicare and Medicaid Services. Retrieved Febuary 2024
from https://www.cms.gov/Medicare/Medicare-Fee-for-ServicePayment/sharedsavingsprogram/Downloads/PioneersMSSPCombinedFastFacts.pdf
CMS. (2017). Medicare Shared Savings Program: Shared Savings and Losses and Assignment
Methodology https://www.cms.gov/Medicare/Medicare-Fee-for-ServicePayment/sharedsavingsprogram/Downloads/Shared-Savings-Losses-Assignment-SpecV5.pdf.
CMS. (2022). Shared Savings Program Fast Facts – As of January 1, 2022. Washington, DC:
Centers for Medicare and Medicaid Services. Retrieved Febuary 2024 from
https://www.cms.gov/files/document/2022-shared-savings-program-fast-facts.pdf
Coe, N. B., White, L., Oney, M., Basu, A., & Larson, E. B. (2023). Public spending on acute and
long-term care for Alzheimer's disease and related dementias. Alzheimer's & Dementia,
19(1), 150-157. https://doi.org/https://doi.org/10.1002/alz.12657

135
Colla, C. H., Lewis, V. A., Kao, L. S., O'Malley, A. J., Chang, C. H., & Fisher, E. S. (2016).
Association Between Medicare Accountable Care Organization Implementation and
Spending Among Clinically Vulnerable Beneficiaries. JAMA Intern Med, 176(8), 1167-
1175. https://doi.org/10.1001/jamainternmed.2016.2827
Colla, C. H., Lewis, V. A., Tierney, E., & Muhlestein, D. B. (2016). Hospitals Participating In
ACOs Tend To Be Large And Urban, Allowing Access To Capital And Data. Health Aff
(Millwood), 35(3), 431-439. https://doi.org/10.1377/hlthaff.2015.0919
Corrigan, P. W., Druss, B. G., & Perlick, D. A. (2014). The impact of mental illness stigma on
seeking and participating in mental health care. Psychological Science in the Public
Interest, 15(2), 37-70.
Daras, L. C., Feng, Z., Wiener, J. M., & Kaganova, Y. (2017). Medicare Expenditures
Associated With Hospital and Emergency Department Use Among Beneficiaries With
Dementia. INQUIRY: The Journal of Health Care Organization, Provision, and
Financing, 54, 0046958017696757. https://doi.org/10.1177/0046958017696757
Desai, U., Kirson, N. Y., Ye, W., Mehta, N. R., Wen, J., & Andrews, J. S. (2019). Trends in
health service use and potentially avoidable hospitalizations before Alzheimer's disease
diagnosis: A matched, retrospective study of US Medicare beneficiaries. Alzheimers
Dement (Amst), 11, 125-135. https://doi.org/10.1016/j.dadm.2018.12.005
Dubois, B., Padovani, A., Scheltens, P., Rossi, A., & Dell'Agnello, G. (2016). Timely Diagnosis
for Alzheimer's Disease: A Literature Review on Benefits and Challenges. J Alzheimers
Dis, 49(3), 617-631. https://doi.org/10.3233/jad-150692
Epstein, A. M., Jha, A. K., Orav, E. J., Liebman, D. L., Audet, A.-M. J., Zezza, M. A., &
Guterman, S. (2014). Analysis of early accountable care organizations defines patient,
structural, cost, and quality-of-care characteristics. Health Affairs, 33(1), 95-102.
Figueroa, J. F., Phelan, J., Orav, E. J., Patel, V., & Jha, A. K. (2020). Association of Mental
Health Disorders With Health Care Spending in the Medicare Population. JAMA Network
Open, 3(3), e201210-e201210. https://doi.org/10.1001/jamanetworkopen.2020.1210
Gozalo, P., Teno, J. M., Mitchell, S. L., Skinner, J., Bynum, J., Tyler, D., & Mor, V. (2011).
End-of-life transitions among nursing home residents with cognitive issues. N Engl J
Med, 365(13), 1212-1221. https://doi.org/10.1056/NEJMsa1100347
Hagi, K., Nosaka, T., Dickinson, D., Lindenmayer, J. P., Lee, J., Friedman, J., Boyer, L., Han,
M., Abdul-Rashid, N. A., & Correll, C. U. (2021). Association between cardiovascular
risk factors and cognitive impairment in people with schizophrenia: a systematic review
and meta-analysis. JAMA psychiatry, 78(5), 510-518.
Happell, B., Ewart, S. B., Bocking, J., Platania‐Phung, C., & Stanton, R. (2016). ‘That red
flag on your file’: misinterpreting physical symptoms as mental illness. Journal of
Clinical Nursing, 25(19-20), 2933-2942.

136
Haye, S., Thunell, J., Joyce, G., Ferido, P., Tysinger, B., Jacobson, M., & Zissimopoulos, J.
(2023). Estimates of diagnosed dementia prevalence and incidence among diverse
beneficiaries in traditional Medicare and Medicare Advantage. Alzheimers Dement
(Amst), 15(3), e12472. https://doi.org/10.1002/dad2.12472
Henderson, C., Noblett, J., Parke, H., Clement, S., Caffrey, A., Gale-Grant, O., Schulze, B.,
Druss, B., & Thornicroft, G. (2014). Mental health-related stigma in health care and
mental health-care settings. The Lancet Psychiatry, 1(6), 467-482.
Hoffman, G. J., Maust, D. T., Harris, M., Ha, J., & Davis, M. A. (2022). Medicare spending
associated with a dementia diagnosis among older adults. Journal of the American
Geriatrics Society, 70(9), 2592-2601. https://doi.org/https://doi.org/10.1111/jgs.17835
Hurd, M. D., Martorell, P., Delavande, A., Mullen, K. J., & Langa, K. M. (2013). Monetary
Costs of Dementia in the United States. New England Journal of Medicine, 368(14),
1326-1334. https://doi.org/10.1056/NEJMsa1204629
Hurd, M. D., Martorell, P., & Langa, K. (2015). Future Monetary Costs of Dementia in the
United States under Alternative Dementia Prevalence Scenarios. J Popul Ageing, 8(1-2),
101-112. https://doi.org/10.1007/s12062-015-9112-4
Jacobson, M., Ferido, P., & Zissimopoulos, J. (2023). Health care utilization before and after a
dementia diagnosis in Medicare Advantage versus traditional Medicare. Journal of the
American Geriatrics Society, n/a(n/a). https://doi.org/https://doi.org/10.1111/jgs.18570
John, A., McGregor, J., Jones, I., Lee, S. C., Walters, J. T., Owen, M. J., O'Donovan, M.,
DelPozo-Banos, M., Berridge, D., & Lloyd, K. (2018). Premature mortality among
people with severe mental illness—New evidence from linked primary care data.
Schizophrenia research, 199, 154-162.
Jones, S., Howard, L., & Thornicroft, G. (2008). 'Diagnostic overshadowing': worse physical
health care for people with mental illness.
Joynt Maddox, K. E. (2018). Financial Incentives and Vulnerable Populations — Will
Alternative Payment Models Help or Hurt? New England Journal of Medicine, 378(11),
977-979. https://doi.org/10.1056/NEJMp1715455
Juster, F. T., & Suzman, R. (1995). An overview of the Health and Retirement Study. Journal of
Human Resources, S7-S56. https://doi.org/10.2307/146277
Knaak, S., Mantler, E., & Szeto, A. (2017). Mental illness-related stigma in healthcare: Barriers
to access and care and evidence-based solutions. Healthcare management forum,
Knox, S., Downer, B., Haas, A., Middleton, A., & Ottenbacher, K. J. (2020). Dementia Severity
Associated With Increased Risk of Potentially Preventable Readmissions During Home
Health Care. Journal of the American Medical Directors Association, 21(4), 519-
524.e513. https://doi.org/https://doi.org/10.1016/j.jamda.2019.09.012

137
Ku, L. J., Pai, M. C., & Shih, P. Y. (2016). Economic Impact of Dementia by Disease Severity:
Exploring the Relationship between Stage of Dementia and Cost of Care in Taiwan. PLoS
One, 11(2), e0148779. https://doi.org/10.1371/journal.pone.0148779
LaMantia, M. A., Lane, K. A., Tu, W., Carnahan, J. L., Messina, F., & Unroe, K. T. (2016).
Patterns of Emergency Department Use Among Long-Stay Nursing Home Residents
With Differing Levels of Dementia Severity. J Am Med Dir Assoc, 17(6), 541-546.
https://doi.org/10.1016/j.jamda.2016.02.011
LaMantia, M. A., Stump, T. E., Messina, F. C., Miller, D. K., & Callahan, C. M. (2016).
Emergency Department Use Among Older Adults With Dementia. Alzheimer Dis Assoc
Disord, 30(1), 35-40. https://doi.org/10.1097/wad.0000000000000118
Langa, K. M., Plassman, B. L., Wallace, R. B., Herzog, A. R., Heeringa, S. G., Ofstedal, M. B.,
Burke, J. R., Fisher, G. G., Fultz, N. H., Hurd, M. D., Potter, G. G., Rodgers, W. L.,
Steffens, D. C., Weir, D. R., & Willis, R. J. (2005). The Aging, Demographics, and
Memory Study: study design and methods. Neuroepidemiology, 25(4), 181-191.
https://doi.org/10.1159/000087448
Lee, D. S., & Lemieux, T. (2010). Regression discontinuity designs in economics. Journal of
economic literature, 48(2), 281-355.
Leicht, H., Heinrich, S., Heider, D., Bachmann, C., Bickel, H., van den Bussche, H., Fuchs, A.,
Luppa, M., Maier, W., Mösch, E., Pentzek, M., Rieder-Heller, S. G., Tebarth, F., Werle,
J., Weyerer, S., Wiese, B., Zimmermann, T., & König, H. H. (2011). Net costs of
dementia by disease stage. Acta Psychiatr Scand, 124(5), 384-395.
https://doi.org/10.1111/j.1600-0447.2011.01741.x
Lewis, V. A., Colla, C. H., Carluzzo, K. L., Kler, S. E., & Fisher, E. S. (2013). Accountable Care
Organizations in the United States: market and demographic factors associated with
formation. Health Serv Res, 48(6 Pt 1), 1840-1858. https://doi.org/10.1111/1475-
6773.12102
Lewis, V. A., Larson, B. K., McClurg, A. B., Boswell, R. G., & Fisher, E. S. (2012). The
promise and peril of accountable care for vulnerable populations: a framework for
overcoming obstacles. Health Aff (Millwood), 31(8), 1777-1785.
https://doi.org/10.1377/hlthaff.2012.0490
Lin, P.-J., Zhong, Y., Fillit, H. M., Chen, E., & Neumann, P. J. (2016). Medicare Expenditures of
Individuals with Alzheimer's Disease and Related Dementias or Mild Cognitive
Impairment Before and After Diagnosis. Journal of the American Geriatrics Society,
64(8), 1549-1557. https://doi.org/https://doi.org/10.1111/jgs.14227
Lin, P. J., Daly, A. T., Olchanski, N., Cohen, J. T., Neumann, P. J., Faul, J. D., Fillit, H. M., &
Freund, K. M. (2021). Dementia Diagnosis Disparities by Race and Ethnicity. Med Care,
59(8), 679-686. https://doi.org/10.1097/mlr.0000000000001577

138
Mauskopf, J., Racketa, J., & Sherrill, E. (2010). Alzheimer's disease: the strength of association
of costs with different measures of disease severity. J Nutr Health Aging, 14(8), 655-663.
https://doi.org/10.1007/s12603-010-0312-6
McGinty, E. E., Baller, J., Azrin, S. T., Juliano-Bult, D., & Daumit, G. L. (2015). Quality of
medical care for persons with serious mental illness: a comprehensive review.
Schizophrenia research, 165(2-3), 227-235.
McWilliams, J. M. (2016). Changes in Medicare Shared Savings Program Savings From 2013 to
2014. Jama, 316(16), 1711-1713. https://doi.org/10.1001/jama.2016.12049
McWilliams, J. M., Chernew, M. E., & Landon, B. E. (2017). Medicare ACO Program Savings
Not Tied To Preventable Hospitalizations Or Concentrated Among High-Risk Patients.
Health Aff (Millwood), 36(12), 2085-2093. https://doi.org/10.1377/hlthaff.2017.0814
McWilliams, J. M., Chernew, M. E., Zaslavsky, A. M., Hamed, P., & Landon, B. E. (2013).
Delivery system integration and health care spending and quality for Medicare
beneficiaries. JAMA internal medicine, 173(15), 1447-1456.
McWilliams, J. M., Gilstrap, L. G., Stevenson, D. G., Chernew, M. E., Huskamp, H. A., &
Grabowski, D. C. (2017). Changes in Postacute Care in the Medicare Shared Savings
Program. JAMA Intern Med, 177(4), 518-526.
https://doi.org/10.1001/jamainternmed.2016.9115
McWilliams, J. M., Hatfield, L. A., Chernew, M. E., Landon, B. E., & Schwartz, A. L. (2016).
Early Performance of Accountable Care Organizations in Medicare. New England
Journal of Medicine, 374(24), 2357-2366. https://doi.org/10.1056/NEJMsa1600142
McWilliams, J. M., Hatfield, L. A., Landon, B. E., & Chernew, M. E. (2020). Savings or
Selection? Initial Spending Reductions in the Medicare Shared Savings Program and
Considerations for Reform. Milbank Q, 98(3), 847-907. https://doi.org/10.1111/1468-
0009.12468
McWilliams, J. M., Hatfield, L. A., Landon, B. E., Hamed, P., & Chernew, M. E. (2018).
Medicare Spending after 3 Years of the Medicare Shared Savings Program. N Engl J
Med, 379(12), 1139-1149. https://doi.org/10.1056/NEJMsa1803388
Michalowsky, B., Flessa, S., Eichler, T., Hertel, J., Dreier, A., Zwingmann, I., Wucherer, D.,
Rau, H., Thyrian, J. R., & Hoffmann, W. (2018). Healthcare utilization and costs in
primary care patients with dementia: baseline results of the DelpHi-trial. The European
Journal of Health Economics, 19, 87-102.
Mitchell, A. J., Vancampfort, D., Sweers, K., van Winkel, R., Yu, W., & De Hert, M. (2013).
Prevalence of metabolic syndrome and metabolic abnormalities in schizophrenia and
related disorders—a systematic review and meta-analysis. Schizophrenia bulletin, 39(2),
306-318.

139
Morris, J. C. (1993). The Clinical Dementia Rating (CDR): current version and scoring rules.
Neurology, 43(11), 2412-2414. https://doi.org/10.1212/wnl.43.11.2412-a
Newcomer, J. W. (2005). Second-generation (atypical) antipsychotics and metabolic effects.
CNS drugs, 19(1), 1-93.
Newcomer, J. W., & Hennekens, C. H. (2007). Severe mental illness and risk of cardiovascular
disease. Jama, 298(15), 1794-1796.
Nyweide, D. J., Lee, W., Cuerdon, T. T., Pham, H. H., Cox, M., Rajkumar, R., & Conway, P. H.
(2015). Association of Pioneer Accountable Care Organizations vs traditional Medicare
fee for service with spending, utilization, and patient experience. Jama, 313(21), 2152-
2161. https://doi.org/10.1001/jama.2015.4930
Oldenburg, C. E., Moscoe, E., & Bärnighausen, T. (2016). Regression Discontinuity for Causal
Effect Estimation in Epidemiology. Curr Epidemiol Rep, 3, 233-241.
https://doi.org/10.1007/s40471-016-0080-x
Olfson, M., Gerhard, T., Huang, C., Crystal, S., & Stroup, T. S. (2015). Premature Mortality
Among Adults With Schizophrenia in the United States. JAMA Psychiatry, 72(12), 1172-
1181. https://doi.org/10.1001/jamapsychiatry.2015.1737
Osborn, D. P., Wright, C. A., Levy, G., King, M. B., Deo, R., & Nazareth, I. (2008). Relative
risk of diabetes, dyslipidaemia, hypertension and the metabolic syndrome in people with
severe mental illnesses: systematic review and metaanalysis. BMC psychiatry, 8(1), 1-14.
Osborn, D. P. J., Levy, G., Nazareth, I., Petersen, I., Islam, A., & King, M. B. (2007). Relative
Risk of Cardiovascular and Cancer Mortality in People With Severe Mental Illness From
the United Kingdom's General Practice Research Database. Archives of General
Psychiatry, 64(2), 242-249. https://doi.org/10.1001/archpsyc.64.2.242
Phelan, E. A., Borson, S., Grothaus, L., Balch, S., & Larson, E. B. (2012). Association of
incident dementia with hospitalizations. Jama, 307(2), 165-172.
https://doi.org/10.1001/jama.2011.1964
Pollack, C. E., & Armstrong, K. (2011). Accountable care organizations and health care
disparities. Jama, 305(16), 1706-1707. https://doi.org/10.1001/jama.2011.533
Rapp, T., Andrieu, S., Molinier, L., Grand, A., Cantet, C., Mullins, C. D., & Vellas, B. (2012).
Exploring the relationship between Alzheimer's disease severity and longitudinal costs.
Value Health, 15(3), 412-419. https://doi.org/10.1016/j.jval.2012.02.003
Rittenhouse, D. R., Shortell, S. M., & Fisher, E. S. (2009). Primary care and accountable care—
two essential elements of delivery-system reform. New England Journal of Medicine,
361(24), 2301-2303.
Robinson, L., Tang, E., & Taylor, J. P. (2015). Dementia: timely diagnosis and early
intervention. Bmj, 350, h3029. https://doi.org/10.1136/bmj.h3029

140
Roehrig, C. (2016). Mental Disorders Top The List Of The Most Costly Conditions In The
United States: $201 Billion. Health Aff (Millwood), 35(6), 1130-1135.
https://doi.org/10.1377/hlthaff.2015.1659
Schwarzkopf, L., Menn, P., Kunz, S., Holle, R., Lauterberg, J., Marx, P., Mehlig, H., Wunder,
S., Leidl, R., Donath, C., & Graessel, E. (2011). Costs of care for dementia patients in
community setting: an analysis for mild and moderate disease stage. Value Health, 14(6),
827-835. https://doi.org/10.1016/j.jval.2011.04.005
St Clair, P., Gaudette, É., Zhao, H., Tysinger, B., Seyedin, R., & Goldman, D. P. (2017). Using
Self-reports or Claims to Assess Disease Prevalence: It's Complicated. Med Care, 55(8),
782-788. https://doi.org/10.1097/mlr.0000000000000753
Stroup, T. S., Olfson, M., Huang, C., Wall, M. M., Goldberg, T., Devanand, D. P., & Gerhard, T.
(2021). Age-specific prevalence and incidence of dementia diagnoses among older US
adults with schizophrenia. JAMA psychiatry, 78(6), 632-641.
Substance Abuse and Mental Health Services Administration (SAMHSA). (2020). 2020 National
Survey on Drug Use and Health. Retrieved July 2022 from
https://www.nimh.nih.gov/health/statistics/mentalillness#:~:text=Prevalence%20of%20Serious%20Mental%20Illness%20(SMI),-
Figure%203%20shows&text=The%20prevalence%20of%20SMI%20was,50%20and%20
older%20(3.4%25).
Thunell, J., Ferido, P., & Zissimopoulos, J. (2019). Measuring Alzheimer's Disease and Other
Dementias in Diverse Populations Using Medicare Claims Data. J Alzheimers Dis, 72(1),
29-33. https://doi.org/10.3233/jad-190310
van Dyck, C. H., Swanson, C. J., Aisen, P., Bateman, R. J., Chen, C., Gee, M., Kanekiyo, M., Li,
D., Reyderman, L., Cohen, S., Froelich, L., Katayama, S., Sabbagh, M., Vellas, B.,
Watson, D., Dhadda, S., Irizarry, M., Kramer, L. D., & Iwatsubo, T. (2023). Lecanemab
in Early Alzheimer's Disease. N Engl J Med, 388(1), 9-21.
https://doi.org/10.1056/NEJMoa2212948
White, L., Fishman, P., Basu, A., Crane, P. K., Larson, E. B., & Coe, N. B. (2019). Medicare
expenditures attributable to dementia. Health Services Research, 54(4), 773-781.
https://doi.org/https://doi.org/10.1111/1475-6773.13134
Wolf, D., Rhein, C., Geschke, K., & Fellgiebel, A. (2019). Preventable hospitalizations among
older patients with cognitive impairments and dementia. Int Psychogeriatr, 31(3), 383-
391. https://doi.org/10.1017/s1041610218000960
Yasaitis, L. C., Pajerowski, W., Polsky, D., & Werner, R. M. (2016). Physicians' Participation In
ACOs Is Lower In Places With Vulnerable Populations Than In More Affluent
Communities. Health Aff (Millwood), 35(8), 1382-1390.
https://doi.org/10.1377/hlthaff.2015.1635

141
Zhu, C. W., Cosentino, S., Ornstein, K., Gu, Y., Andrews, H., & Stern, Y. (2015). Use and cost
of hospitalization in dementia: longitudinal results from a community-based study.
International Journal of Geriatric Psychiatry, 30(8), 833-841.
https://doi.org/https://doi.org/10.1002/gps.4222
Zhu, C. W., Cosentino, S., Ornstein, K., Gu, Y., Scarmeas, N., Andrews, H., & Stern, Y. (2015).
Medicare Utilization and Expenditures Around Incident Dementia in a Multiethnic
Cohort. J Gerontol A Biol Sci Med Sci, 70(11), 1448-1453.
https://doi.org/10.1093/gerona/glv124
Zhu, C. W., Leibman, C., McLaughlin, T., Zbrozek, A. S., Scarmeas, N., Albert, M., Brandt, J.,
Blacker, D., Sano, M., & Stern, Y. (2008). Patient dependence and longitudinal changes
in costs of care in Alzheimer's disease. Dement Geriatr Cogn Disord, 26(5), 416-423.
https://doi.org/10.1159/000164797
Zissimopoulos, J., Crimmins, E., & St Clair, P. (2014). The Value of Delaying Alzheimer's
Disease Onset. Forum Health Econ Policy, 18(1), 25-39. https://doi.org/10.1515/fhep2014-0013

Abstract (if available)

Abstract My dissertation explores two topics in health care of the U.S. aging population, with health care utilization and spending as the underlying theme. The first topic is the Medicare Shared Savings Program (MSSP) implemented by the Centers for Medicare and Medicaid Services. The MSSP creates global incentives for accountable care organizations (ACOs) to reduce health care spending and improve quality of care for assigned Medicare fee-for-service (FFS) beneficiaries. The MSSP is becoming increasingly important in understanding health care use and costs of the aging populations, as a growing share of Medicare FFS beneficiaries are served by providers participating in the program. My dissertation studies how early cohorts of the MSSP affected health care utilization and spending of older U.S. adults with serious mental illnesses and other co-morbidities. My dissertation also addresses the issue of non-random ACO participation to identify the impact of MSSP on health care utilization and spending among beneficiaries seeing more than one provider for primary care services. The second topic of my dissertation is dementia, a common age-related chronic condition that affects 7.6 million older U.S. adults and extracts a heavy financial toll on society. The association of dementia severity at diagnosis with health care use and costs around the time of diagnosis is still not well understood, due in part to data limitations. My dissertation brings together multiple data sources to quantify health care utilization and spending around the time of dementia diagnosis for older adults diagnosed at different stages of dementia, shedding light on how early diagnosis may impact health care use and costs for society. Findings from my dissertation may inform policy interventions that aim at reducing health care spending and improving value of care for the U.S. aging population.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Uncovering hidden figures: disparities in dementia burden in an aging America

PDF

Three essays on health, aging, and retirement

PDF

Thesis on Medicare Part D

PDF

More knowledge, better plans: a study of heterogeneity in dementia prevalence and mid-life cognitive changes

PDF

Residential care in Los Angeles: policy and planning for an aging population

Pdf

Alzheimer's disease: risk factors, value, and alternative payment models

PDF

Essays on work, retirement, and fostering longer working lives

PDF

Racial/ethnic differences in colorectal cancer patient experiences, health care utilization and their association with mortality: findings from the SEER-CAHPS data

PDF

Three essays on estimating the effects of government programs and policies on health care among disadvantaged population

PDF

Essays in opioid use and abuse

PDF

Investigating racial and ethnic disparities in patient experiences with care and health services use following colorectal cancer diagnosis among older adults with comorbid chronic conditions

PDF

Essays on access to prescription drugs

PDF

Topics in U.S. healthcare

PDF

Value in oncology care and opportunities for improvement

PDF

Developing an agent-based simulation model to evaluate competition in private health care markets with an assessment of accountable care organizations

PDF

Adoption and implementation of innovative diagnostic tools for Alzheimer's Disease: challenges and barriers in primary care

PDF

Delivering better care for children with special health care needs: analyses of patient-centered medical home and types of insurance

PDF

Biomarkers of age-related health changes: associations with health outcomes and disparities

PDF

Social engagement and cognitive decline in Mexican Americans: implications for age-friendly cities

PDF

Three essays on aging, wealth, and housing tenure transitions

Asset Metadata

Creator Xu, Shengjia (author)

Core Title Health care utilization and spending of the U.S. aging population

School School of Policy, Planning and Development

Degree Doctor of Philosophy

Degree Program Public Policy and Management

Degree Conferral Date 2024-05

Publication Date 04/04/2024

Defense Date 03/27/2024

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag accountable care organizations,aging populations,alternative payment models,dementia,health care spending,health care utilization,OAI-PMH Harvest

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Chen, Alice (committee chair), Lakdawalla, Darius (committee member), Zissimopoulos, Julie (committee member)

Creator Email xushengj@usc.edu,xushengjia1009@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113871203

Unique identifier UC113871203

Identifier etd-XuShengjia-12762.pdf (filename)

Legacy Identifier etd-XuShengjia-12762

Document Type Dissertation

Format theses (aat)

Rights Xu, Shengjia

Internet Media Type application/pdf

Type texts

Source 20240408-usctheses-batch-1136 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu