Академический Документы
Профессиональный Документы
Культура Документы
Thomas MaCurdy
Jay Bhattacharya
Daniella Perlroth
Jason Shafrin
Anita Au-Yeung
Hani Bashour
Camille Chicklis
Kennan Cronen
Brandy Lipton
Shahin Saneinejad
Elen Shrestha
Sajid Zaidi
Acumen, LLC
Burlingame, CA 94010
[This page is intentionally left blank.]
EXECUTIVE SUMMARY
A large body of research indicates that there exists significant regional variation in health
care spending, utilization, and quality. For instance, the Dartmouth Atlas found that per capita
Medicare reimbursements in Miami were more than twice as high as in Minneapolis.1 Other
studies have also found significant variation in expenditures for end-of-life care and in the
likelihood that individuals are diagnosed with a specific disease.2 In the popular media, Atul
Gawandes article in the New Yorker magazine further advanced the notion that variation in
physicians chosen practice patterns drives variation in Medicare costs observed even in cities
close to one another.3 Gawande uses the Texan cities of McAllen and El Paso to highlight this
point. By first identifying the source of this geographic variation, policymakers can potentially
develop and implement initiatives to alter practice patterns in high-cost areas.
This study contributes to this debate by examining geographic variation in the volume
and intensity of per capita health care services and spending for both Medicare and Medicaid
beneficiaries. This project was undertaken under the direction of the Institute of Medicine
(IOM) and includes the analyses performed at their behest. This study does not necessarily
reflect Acumens own analytic approach but rather was designed to follow specifications
required by IOM to promote consistency across contractors. This report therefore presents a
comprehensive set of results that examine myriad issues underlying regional differences in
Medicare and Medicaid spending. Specifically, this report aims to answer the following research
questions:
1. How much geographic variation exists in per capita volume of healthcare services?
2. Are regions with high utilization levels likely to have high utilization rates in the
future?
3. Is the variation in the volume of medical services greater within or across regions?
4. Do regions that provide a high volume of medical services when treating beneficiaries
for a given disease also provide a high volume of medical services when treating all
other diseases?
5. What types of services are the primary drivers of regional variation in the utilization
of medical services?
6. Are areas with high utilization levels more likely to have high quality care?
1
The Center for the Evaluative Clinical Services and Dartmouth Medical School, The Dartmouth Atlas of Health
(Chicago: American Hospital Publishing, Inc., 1996).
2
Y Song et al., "Regional Variations in Diagnostic Practices," New England Journal of Medicine 2010, no. 363
(2010).
3
Atul Gawande, "The Cost Conundrum: What a Texas town can teach us about health care," New Yorker(June
2009), http://www.newyorker.com/reporting/2009/06/01/090601fa_fact_gawande.
Methodological Approach
This study measures each regions average per capita Medicare and Medicaid
expenditures and utilization of medical services between 2007 and 2009. Calculating these
values requires four steps:
To quantify the amount of geographic variation in the provision of health care services,
this study first examines regional variation in both health care expenditures and utilization.
Expendituresmeasured on a per capita basisinclude cost incurred by both Medicare and the
beneficiary. To evaluate regional differences in utilization patterns, this study measures regional
variation in price-standardized spending levels. Input price standardization removes geographic
variation in the price of inputs, such as labor costs, and allows for an assessment of changing
patterns of utilization overall. In addition to analysis of the overall input price adjusted
expenditures, this study also analyzes the input price adjusted expenditures stratified by seven
medical service categories, such as acute hospitalization or diagnostic services, to determine if
regional variation overall is similar to regional variation for specific types of services.
After defining the spending and utilization measures of interest, the second step of the
methodology next defines the beneficiary of interest as both an aggregate cohort and cohorts of
beneficiaries with specific health conditions. The aggregate cohort includes all Medicare
beneficiaries that enrolled in Medicare fee-for-services (FFS). In addition to the aggregate
cohort, this study includes 15 condition cohorts selected by the Institute of Medicine (IOM)
Committee based on a variety of factors including disease prevalence, disease incidence, costs of
ii Acumen, LLC
treatment, and the likelihood of regional variation in the course of treatment due to variation in
demand-side or supply-side factors. The chosen conditions include both acute and chronic
conditions and three incident cancers. The acute conditions include acute/ischemic stroke, acute
myocardial infarction (AMI), pneumonia, cataracts, and cholecystectomy. The acute condition
cohorts cholecystectomy and cataracts are defined, in part, using procedures. The chronic
conditions include diabetes, rheumatoid arthritis, depression, congestive heart failure (CHF),
coronary heart disease (CHD), chronic obstructive pulmonary disease (COPD), and low back
pain. The three cancer cohortsbreast cancer, lung cancer, and prostate cancerinclude
beneficiaries newly diagnosed with cancer.
To measure regional variation in spending and utilization levels for each of these cohorts,
the third step defines a region as a hospital referral region (HRR). The Dartmouth Atlas
defines 306 HRRs based on referrals for surgical procedures and neurosurgery. Each HRR
contains at least one city where both major cardiovascular surgical procedures and neurosurgery
are available. To assign beneficiaries to regions, the analysis uses the beneficiary ZIP code at the
beginning of a given year. If the beneficiarys ZIP code changes during an episode, all claims are
assigned to the ZIP code at the beginning of the episode because moving is considered to be
within the set of treatment options for beneficiaries in the original ZIP code. Beneficiary-level
data is aggregated to the ZIP code level and then to the HRR level. Regional variation in
expenditures and utilization is also examined by hospital service area (HSA) and metropolitan
statistical area (MSA); however, these results are not included in this report.
Although many of the analyses in this study use the first three steps to create measures of
geographic variation in spending and utilization, in the fourth and final step this study applies a
risk adjustment methodology to control for regional differences in patient case mix. The risk
adjustment framework accounts for differences in beneficiary demographics and severity of
illness andin certain specificationsmarket-level factors as well. The general risk adjustment
methodology uses a linear ordinary least squares regression model to predict the value of the
outcome variable given a set of observable beneficiary characteristics. This study gathers
beneficiary-level information from Medicare claims to account for patient case mix. The risk
adjustment model includes the following independent variables: beneficiary age, sex, health
status (HCCs), income/pharmacy benefit, partial-year enrollment, the interaction between age
and sex, a new enrollee indicator, and an indicator for the year of analysis. The region-level
value of the risk-adjusted outcome value equals the average difference between the observed
levels of the outcome value and the values predicted by the risk adjustment model.
iv Acumen, LLC
Figure 2: HRR Utilization Rank 2007-2008, Medicaid
Within-HRR variation in spending and utilization is significantly larger than the across-
HRR variation for both Medicare and Medicaid. The average standard deviation of per-capita
utilization for beneficiaries within each HRR is $1,621, whereas the standard deviation of
average Medicare utilization levels across HRRs is $84. In other words, variation in utilization
levels for beneficiaries within an HRR is over nineteen times greater than the variation in
utilization for the average patient across HRRs. The findings from the Medicaid analysis are
comparable. The average standard deviation within HRRs of price-standardized risk-adjusted
Medicaid expenditures is $2,446, whereas the standard deviation of average Medicaid costs
across HRRs is $334. For Medicaid, variation in patient utilization for patients within an HRR is
over seven times greater than the variation in utilization for the average patient across HRRs.
Variability both within- and across-HRR is larger for an HRRs Medicaid per capita utilization
figures than the corresponding Medicare per capita utilization figures.
High Cost Users are the Key Drivers of Regional Variation in Spending
This report presents two pieces of evidence indicating that regional variation in spending
is due primarily to the relative amount of resources used to treat the most expensive beneficiaries
rather than the cost of treating a typical beneficiary. The first piece of evidence is that
Medicare and Medicaid cost distributions are highly right-skewed. The median Medicare cost
($310) is far below the mean cost ($964). The right-skewed nature of the Medicare per capita
spending level persists even after removing variation due to regional differences in prices, patient
demographics, and observed beneficiary severity of illness. For the aggregate Medicare cohort,
Not only are median costs significantly below mean costs for both Medicare and
Medicaid, but an HRRs average per capita utilization levels are much more highly correlated
with the utilization levels of high-cost beneficiaries in that HRR than the utilization of the typical
beneficiary in that HRR. For the period 2007 to 2009, Figures 5 and 6 plot each HRRs rank by
mean utilization level against its rank by median utilization levels for Medicare and Medicaid
respectively; Figure 7 and Figure 8 plot HRR rank by mean utilization level versus its rank by its
resource use intensity for treating beneficiaries at the 90th percentile. Although there is a
moderate level of correlation between an HRRs mean and median Medicare utilization levels,
the relationship is weak in general. The relationship between the mean and the 90th percentile
rank, however, is much stronger. The same qualitative relationship exists for the Medicaid
analysis as well. These figures illustrate that stability of average HRR utilization levels across
years is due to the persistence in the cost of treating the highest-cost beneficiaries rather than the
cost of treating beneficiaries at the median.
vi Acumen, LLC
Figure 4: Medicaid Utilization Levels by HRR, Mean vs. Median
The use of post-acute care services is the main driver of HRR-level variation in
utilization levels for both Medicare and Medicaid. To demonstrate that this is the case, Figures 9
and 10 plot total cost residuals against post-acute and acute care (i.e., inpatient) residuals for
Medicare and Medicaid, respectively. These figures calculate the residual utilization levels as the
difference between the observed per capita utilization on each service category and the expected
per capita utilization levels based on patient case mix. A positive residual for a given service
indicates that utilization of that service is higher than expected based on patient case mix, while a
negative residual indicates that utilization of that service is lower than expected. From left to
right on the x-axis, the HRRs are ordered from lowest to highest (risk-adjusted) utilization levels.
Based on these graphs, one can clearly see that HRRs that have the lowest total cost
residuals (on the left) are more likely to have the lowest post-acute cost residuals, and HRRs that
have the highest total cost residuals (on the right) are more likely to have the highest post-acute
cost residuals. The relationship between overall spending and acute care services is not only
much weaker, but the residuals are of a smaller magnitude. Figure 10 shows that post-acute
services are also the main driver of regional variation in Medicaid utilization levels. Additional
analysis indicates that the use of home health services is the key driver of regional variation in
Medicare post-acute spending.
This report also examines regional variation in Medicare expenditures for beneficiaries
enrolled in Medicare Advantage (MA). The following summarizes the main findings of the
Medicare Advantage analysis:
Average per capita MA expenditures are higher than average per capita FFS
expenditures ($986 and $958, respectively).
MA expenditures per person are less variable than FFS expenditures (with standard
deviations of $414 and $1,695, respectively).
Regions that are high- or low-cost in one year tend to be similarly high- or low-cost in
the next. The correlation between an HRRs per capita MA expenditures in 2007 and
its per capita MA expenditures in 2009 is 0.94.
HRRs with more MA plan competition do not necessarily have lower MA premiums.
The correlation between an HRRs per capita FFS and MA expenditures is 0.66
x Acumen, LLC
Figure 10: Positive Correlation between Medicare Utilization and Avoidable Complications
Regions that exhibit high utilization of Medicare services do not necessarily have high
utilization levels of Medicaid services. Using the price-standardized, risk adjusted cost from
2007 through 2009, Figure 13 presents each HRRs per capita resource use level rankings for
Medicare beneficiaries on the horizontal axis and resource use ranks for Medicaid beneficiaries
on the vertical axis.4 The graph shows almost no relationship between Medicare and Medicaid
utilization levels. In fact, the correlation between an HRRs Medicare and Medicaid resource
use is -0.07, indicating that HRRs with high per capita Medicare utilization levels are no more
likely to have high per capita Medicaid utilization levels than any other HRR.
4
The size of the bubble that represents each HRR is proportional to the number of Medicaid episodes in the
aggregate cohort in that HRR from 2007 through 2009.
Summary of Findings
This report presents several interesting findings regarding regional variation in Medicare
and Medicaid spending and utilization. Variation in spending and utilization across regions is
large, but variation among beneficiaries within each HRR is substantially larger than variation
across HRRs. For both Medicare FFS and Medicaid beneficiaries, an HRRs average per capita
resource utilization level is relatively stable over time; high utilization regions tend to remain so
across years. This strong year-to-year correlation, however, is not driven by similar treatment
patterns for the typical patient over time; rather the correlation is due in large part to
persistence in the cost of treating the highest-cost beneficiaries over time. Further, areas with
high resource intensity overall do not necessarily use more of all types of services. Despite this
fact, the use of post-acute care is a key predictor of HRR utilization levels for both the Medicare
FFS and Medicaid analysis. Finally, there exists practically no relationship between the
utilization levels of an HRRs Medicare FFS beneficiaries and the utilization levels of an HRRs
Medicaid beneficiaries. Table 1 presents some key statistics that support the major findings of
this report.
Table 1: Regional Variation in Spending, Utilization and Quality, Medicare vs. Medicaid
Research
Question Research Question Description Medicare Findings Medicaid Findings
Number
Variation in Spending Across All Average Cost : $958 Average Cost: $1,094
1
Beneficiaries Standard Deviation: $1,695 Standard Deviation: $2,767
Correlation: HRR Utilization
2 0.97 0.90
Levels Over Time (2007 - 2008)
5
Positive numbers indicate higher spending is correlated with higher quality
6
The Center for the Evaluative Clinical Services and School, The Dartmouth Atlas of Health.
7
Song et al., "Regional Variations in Diagnostic Practices."
8
Gawande, "The Cost Conundrum: What a Texas town can teach us about health care".
9
Medicare Payment Advisory Commission, "Measuring Regional Variation in Service Use: Report to Congress,"
(December 2009), http://www.medpac.gov/documents/Dec09_RegionalVariation_report.pdf.
10
Medicare FFS includes Medicare Parts A and B.
11
J. D. Reschovsky et al., "Following the Money: Factors Associated with the Cost of Treating High-Cost Medicare
Beneficiaries," Health Services Research no. doi: 10.1111/j.1475-6773.2011.01242.x.
12
Luisa Franzini, Osama I. Mikhail, and Jonathan S. Skinner, "McAllen And El Paso Revisited: Medicare
Variations Not Always Reflected In The Under-Sixty-Five Population," Health Affairs 29, no. doi:
10.1377/hlthaff.2010.0492 (2010).
13
Darius Lakdawalla, Tomas Philipson, and Dana Coldman, "Addressing Geographic Variation and Health Care
Efficiency: Lessons for Medicare from Private Health Insurance," AEI Health Policy Outlook 2, no. July (2010),
http://www.aei.org/docLib/2010-7-No-2-g.pdf.
1. How much geographic variation exists in per capita volume of healthcare services?
2. Are regions with high utilization levels likely to have high utilization rates in
subsequent years?
3. Is the variation in the volume of medical services greater within or across regions?
4. Do regions that provide a high volume of medical services when treating beneficiaries
for a given disease also provide a high volume of medical services when treating all
other diseases?
5. What types of services are the primary drivers of regional variation in the utilization
of medical services?
6. Are areas with high utilization levels more likely to have high quality care?
7. Are regions with high utilization levels in Medicare likely to have high utilization
levels in Medicaid?
To answer these questions, this report relies on claims data covering the universe of
Medicare and Medicaid FFS beneficiaries. Sections 2 through 4 describe how this study uses
these data to answer the questions above. Section 2 identifies the data sources used for the
analysis and the beneficiaries that are included in the aggregate cohort and 15 condition-specific
cohorts. Section 3 defines the three outcome measures. Section 4 describes the methodology,
including the risk adjustment regression specifications and the geographic region definitions.
In addition to producing a wealth of statistics that researchers can use for future studies,
this report also directly evaluates the six research questions posed above. Section 5 addresses
each of these questions for Medicare FFS beneficiaries, and Section 6 answers these questions
for Medicaid beneficiaries. The latter section also measures the correlation between an HRRs
average utilization of Medicare services and an HRRs average utilization of Medicaid services.
Section 7 presents a summary of this reports major findings. In addition to investigating
regional variation in medical spending and utilization for Medicare and Medicaid FFS
2 Acumen, LLC
beneficiaries, Appendix F measures regional variation in spending for Medicare Advantage
beneficiaries.
This section describes the methodology used to transform Medicare and Medicaid claims
data into useful analytical files for the aggregate and fifteen condition cohorts. This section
proceeds in two parts. Sections 2.1 and 2.2 define the data sources, exclusions, and the
population for the Medicare and Medicaid analyses, respectively. Section 2.3 describes the steps
the analysis uses to define the aggregate and 15 condition cohorts.
To synthesize all of the medical events that comprise each Medicare beneficiarys
expenditure levels, treatments, and health history, this analysis relies on the universe of Medicare
claims data from 2005 through 2010. Although the years of analysis are 2007 through 2009, this
study uses data from 2005 through 2006 to capture additional beneficiary health history
information. This study uses 2010 data to capture claims information for episodes beginning in
2009 that end in 2010. The remainder of this section contains three parts. Section 2.1.1 presents
the Medicare data sources; Section 2.1.2 defines the exclusions applied to the Medicare
population; and Section 2.1.3 discusses the Medicare population included in the analysis.
2.1.1 Medicare Data Sources
The Medicare investigation uses the claims, enrollment, and assessment data sources
listed in Table 2.1 to produce relevant analytical files. Medicare Part A, B, and D claims files are
episodic, rather than longitudinal, data files where observations occur when Medicare
beneficiaries interact with providers who are to be paid by Medicare. The data on the claims
describe the cost for services rendered, what services were provided during the interaction, who
4 Acumen, LLC
provided these services, and a wide range of information regarding the beneficiarys
demographics and health condition. Next, Medicare Part A, B, C, and D enrollment data files are
historical with an observation every time a beneficiarys status changes. These files contain
detailed data on all individuals entitled to Medicare, including demographic information,
enrollment dates, third party buy-in information, and Medicare managed care (MC) enrollment.
Table 2.1: Medicare Data Sources
Data Source Years Data Files
Common Working Files (CWF) for Home Health (HH), Physician (PB),
Medicare Parts A and B 2005 -
Inpatient (IP), Skilled Nursing Facility (SNF), Outpatient (OP), Hospice
Claims 2010
(HS), and Durable Medical Equipment (DME) claims
2006 -
Medicare Part D Claims Prescription Drug Event (PDE)
2009
Enrollment Database (EDB)
Medicare Part A, B, and 2005 -
Common Medicare Environment (CME)
C Enrollment Data 2010
Enterprise Cross-Reference (ECR) Files
MARx files: Full Enrollment Files, Monthly Membership Files, Risk Scores
Medicare Part C and D 2005 -
Enrollment Data 2010 Risk Adjustment Processing System (RAPS)
HPMS Files: Beneficiary Cost, Formulary and Pharmacy Files for Part D
The Medicare analysis includes only claims for FFS beneficiaries in the months that they
are enrolled. To limit the sample in this manner, this study excludes beneficiaries only enrolled
in Medicare Advantage (also known as Part C or Medicare Managed Care) throughout the
observation period from the analysis because information on utilization of and payment for
medical services rendered is not available in the MA data.14 A Medicare beneficiary is
considered to be enrolled in FFS for a given month if the Medicare Enrollment Database shows
enrollment in Part A or B and not in Part C during that month. When a beneficiary switches
between Medicare FFS and Medicare Advantage, only the months with FFS enrollment are
included in the episode. Months where a beneficiary is not enrolled in FFS but has FFS Part D
claims are excluded from the observation window as the beneficiary is not enrolled in Medicare
FFS Parts A or B. For the condition cohorts, beneficiaries must be enrolled in Medicare Part A or
Part B during the first month of the observation window. In addition to excluding episodes with
third party payer costs on claims, the Medicare analysis also excludes episodes with the
beneficiary listed as having a third-party primary payer in the observation window in the EDB.
Appendix A.5 illustrates the number and percent of Medicare beneficiaries who are excluded
from the analysis due to each restriction criteria.
14
Beginning in 2012, CMS will collect encounter claims data from MA plans. These data, however, are not
available for this studys time frame.
Because Medicare eligibility begins at age 65 for those without disabilities, the majority
of Medicare beneficiaries are age 65 and older. Table 2.2 shows the number of episodes by sex,
race, dual-eligibility status, episode end, and age in the aggregate 2007 through 2009 analysis. In
total, 78 percent of episodes are for beneficiaries who are at least 65. The majority of Medicare
beneficiaries are not eligible for Medicaid (83 percent), and most episodes do not end in death
(96 percent).
15
A very small number (less than 0.00001 percent) of beneficiaries who are missing sex or date of birth in the
enrollment files (EDB/EL) due to coding errors are excluded from this analysis.
16
Stephanie E. Anthony et al., "Medicaid Managed Care for Dual Eligibles: State Profiles," The Kaiser Commission
on Medicaid and the Uninsured 14(2000),
http://www.kff.org/medicaid/loader.cfm?url=/commonspot/security/getfile.cfm&PageID=13759.
6 Acumen, LLC
Table 2.2: Medicare Demographics
# Episodes
Beneficiary Category Percent
(millions)
All Beneficiaries 104.7 100%
Female 58.2 56%
Male 46.6 44%
White 87.5 84%
Black 10.7 10%
Asian 2.0 2%
Hispanic 2.3 2%
Other 2.1 2%
Unknown 0.2 <1%
Dual-Eligible 21.7 21%
Not Dual-Eligible 83.0 79%
Living during Entire
100.1 96%
Episode
Not Living during Entire
4.6 4%
Episode
Age under 65 23.4 22%
Age 65-69 22.7 22%
Age 70-74 18.3 17%
Age 75-79 15.7 15%
Age 80-84 12.6 12%
Age 85-89 7.7 7%
Age 90 and over 4.4 4%
Paralleling the Medicare analysis, the Medicaid analysis relies on the universe of
Medicaid claims data from 2005 through 2010. Medicaid is a publicly-financed health care
coverage program for low-income people. Like the Medicare analysis, this study uses data from
2005 through 2006 to capture additional beneficiary health history information and uses 2010
data to capture claims information for episodes beginning in 2009 that end in 2010. The
remainder of this section contains three parts. Section 2.2.1 presents the Medicaid data sources,
and Section 2.2.2 defines the exclusion applied to the Medicaid population. Section 2.3.2
describes the Medicaid population included in the analysis.
Medicaid data come in two basic forms: as MAX (Medicaid Analytic eXtract) files, and
as elements from MSIS (Medicaid Statistical Information System). MAX data is a subset of
MSIS data that is produced after verification and enhancement of the MSIS files and arranged by
beneficiary in an annual time frame. The MSIS data, on the other hand, comprise raw claims and
beneficiary enrollment information uploaded from states on a quarterly/monthly basis. The
MAX/MSIS data consist of five types of files: personal summary (PS), inpatient hospital (IP),
long-term care (LT), other services/therapy (OT), and prescription drugs (RX). While 1999 to
2009 MAX files are currently available for all but a few small states, files for year of service
2010 are only being released now. In contrast, MSIS contains the most recent data submitted by
each state, with updates usually done at least quarterly. Since claims in MSIS come from states
exercising different submission protocols, adjustments must be made for some data elements to
create common formats and definitions. When combined, MAX and MSIS supply a
comprehensive set of Medicaid enrollment and claims information covering 1999 to the present
time. Table 2.3 presents the Medicaid data sources and Medicare equivalents.
Table 2.3: Medicaid Data Sources and Medicare Equivalents
Medicaid Data Files Medicare Equivalent(s)
Inpatient stays (IP) IP
Long-term care stays (LT) SNF
Other therapy (OT) HH, DME, PB, OP
Prescription drugs (RX) PDE
Eligibility records (EL) (MSIS) EDB
This report relies on both MAX and MSIS data. For claims information in 2009 or
before, MAX data is used. There are 3 states, however, without complete MAX information at
the time of the publication of this report. States with limited coverage in the 2009 MAX data
include: Idaho, Maine, Missouri and Wisconsin. For these states, we supplemented the MAX
data with MSIS claims information.17 Additionally, for the condition cohorts, episodes that
begin in 2009 can spill over into 2010. For instance, a 12 month episode that begins on
December 1, 2009 would include all claims through November 30, 2010. Because MAX data
was not currently available at the time of this reports publication, this report relies exclusively
on MSIS for all 2010 information.
Working with Medicaid data presents a number of unique challenges compared to the
Medicare data. Although most Medicaid states programs now use HCPCS/CPT codes to report
procedures and equipment in OT, ICD-9 procedure codes in IP claims and report prescription
drug use in the RX file using 11-digit NDC codes, states deviate from national coding standards,
17
MSIS data was used for: Idaho (2009 for all file types), Maine (2005-2009 for the IP, LT and OT files), Missouri
(2009 for RX file type), and Wisconsin for 2009 for all file types.
8 Acumen, LLC
for instance, by reporting local provider identification numbers or DEA numbers to identify
service providers and insurers in place of NPI. Also, several states, including Maine and New
York, report a significant number of procedures on their claims using unique local coding
systems rather than HCPCS/CPT codes. Thus, identifying providers and procedures in Medicaid
claims is complicated by the lack of unified, nationwide databases containing all local Medicaid
provider identification numbers or local procedure codes.
Like the Medicare study, the Medicaid analysis applies several eligibility restrictions to
require complete data for each beneficiary.19 The Medicaid analysis also excludes beneficiaries
with missing ZIP codes; with third party payer costs; and who are simultaneously enrolled in
18
"2007 State Perspectives Medicaid Pharmacy Policies and Practices," National Association of State Medicaid
Directors.
19
A very small number (less than 0.00001 percent) of beneficiaries are missing sex or date of birth in the enrollment
files (EDB/EL) due to coding errors and are excluded from this analysis.
Second, beneficiaries who are not covered for the full set of Medicaid services are
excluded because their claims record likely contains an incomplete picture of their healthcare
spending and utilization levels. The Medicaid data identifies these beneficiaries using a benefit
restrictions indicator variable. For example, the Family Planning Access Care and Treatment
(PACT) waiver in California provided comprehensive family planning services through
20
Daniel R. Levinson, Inspector General, "Medicaid Managed Care Encounter Data: Collection and Use," 12, no.
OEI-04-07-00240 (2009), http://oig.hhs.gov/oei/reports/oei-07-06-00540.pdf.
21
"Medicaid - A Primer: Key information on Our Nations Health Coverage Program for Low-Income People," The
Kaiser Commission on Medicaid and the Uninsured, http://www.kff.org/medicaid/upload/7334-04.pdf.
10 Acumen, LLC
Medicaid to 2.6 million beneficiaries in 2009 not otherwise eligible for Medicaid. Because these
beneficiaries only received coverage for family planning services through Medicaid, they are
indicated with benefit restriction flags and excluded from the analysis as each may have an
incomplete health history. Third, the Medicaid study excludes all dual-enrolled beneficiaries
because their costs are first covered by Medicare then by Medicaid. Thus, to avoid double-
counting, only the Medicare analysis includes dual-enrolled beneficiaries.
2.2.3 Medicaid Population Summary Statistics
Different eligibility groups can receive different Medicaid benefits within a state or
county, but there are nationally mandated minimums that mainly cover limited-income families
with children and pregnant women.22 For example, infants born to Medicaid-eligible women as
well as pregnant women whose family income is at or below 133 percent of the federal poverty
line are covered by Medicare. As a result, the Medicaid population is much younger and includes
a larger share of females than the Medicare population. Table 2.6 below presents the number of
Medicaid episodes in the 2007 through 2009 analysis by sex, race, and age.
22
Ibid.
23
"State Health Facts," Kaiser Family Foundation, http://www.statehealthfacts.org/.
24
"Medicaid Managed Care: Key Data, Trends, and Issues," Kaiser Commission on Medicaid and the Uninsured,
http://www.kff.org/medicaid/upload/8046-02.pdf.
This report examines regional variation in cost, utilization, and quality for the aggregate
Medicare and Medicaid cohorts and for 15 condition cohorts. The aggregate cohort includes all
beneficiaries and counts those beneficiarys claims during the months they are enrolled over the
observation period, which is a calendar year. The condition cohorts include beneficiaries with
certain health conditions and count their claims during the months they are enrolled. The chosen
conditions include both acute and chronic conditions and three incident cancers. The acute
conditions include acute/ischemic stroke, acute myocardial infarction (AMI), pneumonia,
cataracts, and cholecystectomy. The acute condition cohorts cholecystectomy and cataracts are
defined, in part, using procedures. The chronic conditions include diabetes, rheumatoid arthritis,
depression, congestive heart failure (CHF), coronary heart disease (CHD), chronic obstructive
12 Acumen, LLC
pulmonary disease (COPD), and low back pain. The three incident cancer cohorts include breast
cancer, lung cancer, and prostate cancer.
Each condition cohort algorithm follows a three step procedure. These steps include:
2. Identifying the start of an episode using a clean period requirement (if any), and
Sections 2.3.1 through 2.3.3 describe each of these three steps in detail. Section 2.3.4 discusses
challenges to the cohort-based approach, and outlines the restrictions that are applied to both the
aggregate and condition cohorts.
This study defines beneficiaries as members of a given condition cohort using diagnosis,
procedure, and prescription drug codes. Medical experts on the IOM Committee selected the
codes deemed most predictive of the conditions represented by each of the twelve chronic and
acute condition cohorts using a consensus process. This study, in consultation with IOM and
Peter Bach of the Sloan Kettering Cancer Center, selected the diagnosis, procedure, and
prescription drug codes to predict membership in the three incident cancer cohorts using an
empirical approach based on the academic literature. Appendix A.1 classifies the cohorts as
acute or chronic and gives their clean period requirements, which are discussed below. Appendix
A.2 provides a full listing of the condition cohort algorithms, and Appendix A.3 has a full listing
of the condition cohort codes. Appendix A.4 describes the methodology used to determine the
cancer condition cohort specifications.
To reduce the number of false-positives included in the condition cohorts, the analysis
applies a number of restrictions to refine the cohort definitions. Diagnosis and procedure codes
on claims made up entirely of laboratory services are not eligible to qualify a beneficiary for a
cohort.25 This restriction aims to reduce the likelihood that rule-out diagnoses affect
beneficiary health status measures. Rule-out diagnoses occur when providers indicate whether
the beneficiary received a test for a given disease rather than indicating whether he or she has the
given condition. The analysis also excludes claims reporting zero Medicare payments. If no
payment was made, it is unlikely that a service was rendered.26 The analysis also excludes
25
Lab codes are defined using HCPCS codes. HCPCS codes that begin with EKABCLV, codes that have a
BETOS code that begins with I or T, codes that are in the range 70010-76999 or 78000-78999, and codes that are
included in the 2010 clinical lab fee schedule are included in the labs category.
26
PB claims must have a valid line item where a valid line item has a pricing indicator of A, R, or S indicating the
claim was paid.
To identify the beginning of a new episode of care, the cohort definitions for the acute
conditions acute/ischemic stroke, acute myocardial infarction, and pneumonia and all
incident cancers breast cancer, prostate cancer, and lung cancer require a clean period.
During the clean period, claims must not contain diagnosis or procedure codes related to the
condition in question, as determined in step 1. If a beneficiarys claims submitted during the
clean period do contain a combination of codes determined in step 1, this suggests the
beneficiary developed the condition prior to the index date, the date of the first claim
containing a condition-relevant code during the period of analysis. Since this study aims to
analyze the costs of incident acute conditions, beneficiaries with claims violating the clean
period requirement are excluded from the acute condition cohorts and the incident cancer
cohorts. For example, if claims with diagnosis, procedure, or prescription drug codes associated
with acute myocardial infarction (AMI) are observed for a given beneficiary on March 1, 2008
and March 20, 2008, the clean period approach assumes that these claims are related to the same
episode of care since the AMI clean period is 60 days. The clean period requirement ensures
these claims are grouped together in the same episode of care if both claims are submitted during
the period of analysis. If claims for AMI are observed for a given beneficiary on March 1, 2008
and subsequently on August 1, 2008, it is assumed that these claims are related to separate
episodes of care and the clean period requirement ensures they are analyzed separately.
Appendix A.1 lists the clean period requirement, if any, for each cohort. For cohorts with clean
periods, the Medicare analysis does not require that beneficiaries be continuously enrolled during
the clean period to be eligible for the cohort. The Medicaid analysis, however, does require
continuous enrollment during the clean period to ensure that the clean period is valid.
27
The physician specialty field must contain a code that is eligible for risk adjustment: "Acceptable Physician
Specialty Types for Risk Adjustment Data Submission," Centers for Medicare & Medicaid Services,
http://www.csscoperations.com/internet/Cssc.nsf/files/Risk-Adjust-Physn-Spec-
Types031511_040811.pdf/$FIle/Risk-Adjust-Physn-Spec-Types031511_040811.pdf.
14 Acumen, LLC
There are several advantages to the use of a clean period requirement. First, the clean
period requirement is straightforward to implement using the Medicare claims data.
Implementing the clean period approach requires a list of diagnostic and procedural codes for
inclusion in the cohort and a choice of the length of the clean period for each cohort. For this
project, the IOM Committee selected the length of the clean period using a consensus-based
approach. In addition, the use of a clean period has widespread use in both commercial and
governmental applications. Commercial grouping software developed by Ingenix and Thomson
Reuters uses a clean period as does the medical spending per beneficiary measure CMS uses
within Hospital Compare.28
Although imposing a clean period restriction has several advantages, meeting this
requirement is imperfectly correlated with having an incident rather than prevalent condition.
Beneficiaries with no evidence of a given condition during the clean period may have a prevalent
condition with past treatment of the condition occurring outside of the clean period or they may
simply have not sought treatment for their condition. In this case, the methodology incorrectly
assigns these beneficiaries to the given condition cohort. Beneficiaries with evidence of a given
condition during the clean period may have multiple incident cases of that condition, each
requiring its own episode of care. In this case, the methodology incorrectly eliminates these
beneficiary episodes from the given condition cohort.
To determine the beginning of a new episode of care for cohorts without a clean period
requirement, the date of the first claim containing a condition-relevant code is used as the index
date. Clean periods are not required for the acute condition cohorts cholecystectomy and
cataracts because these cohorts represent procedures, and the majority of costs related to these
conditions will be incurred in the immediate period before (pre-operative costs) and after (post-
operative costs) the procedure. For example, the costs related to cataracts surgery include a
physicians consultation and examination prior to surgery, the costs of the surgery, prescription
eye drops to prevent infection after surgery, physician visits post-surgery to monitor
complications such as infection, persistent inflammation, or changes in eye pressure, and the
costs of treating any complications that may occur. In addition, because chronic conditions recur
and subside on a frequent basis, a clean period is not required for the chronic condition cohorts.
Once the start of an episode is identified, this study implements an observation period for
each condition cohort to capture the relevant costs associated with an episode of care. The
observation period is the length of the period of analysis, which begins on the index date. IOM
28
T. MaCurdy et al., "Optimal Pay-for-Performance Scores: How to Incentivize Physicians to Behave Efficiently,"
http://www.cms.gov/reports/downloads/MaCurdy_Incentivize_Physicians_Optimal_P4P_Scores_Feb_2011.pdf.
Conducting the analysis at the condition cohort level allows for examination of the
regional variation in costs, utilization, and quality for individuals with similar health profiles;
however, this approach presents several challenges. First, diagnosis and procedure codes may be
reported differently across regions. The same patient may receive a different diagnosis
depending on which physician he or she sees, and if physicians in certain regions diagnose
patients in systematically different ways than physicians in other regions, the results of a cohort-
based analysis of regional variation in health care may be biased. For example, if upcoding
diabetes diagnoses is common in the Northeast, then beneficiaries included in the diabetes cohort
residing in the Northeast will be relatively healthy since some of these individuals may not meet
the criteria for a diabetes diagnosis in other regions. If upcoding occurs, these beneficiaries
costs and utilization will be relatively low compared to other regions not due to treatment choice
but due to the cohort to which this study classifies them. Second, regions may also differ in the
propensity to screen for certain conditions. If some regions are more likely to identify certain
conditions at an earlier stage, these conditions may be less expensive to treat and treatment may
29
According to a study published in the Journal of General Internal Medicine, while the majority of patients rated
symptoms of fatigue (79%) and cough (80%) as moderate to severe at the time of diagnosis, a minority of patients
rated symptoms as moderate or severe at the day 30 or day 90 follow-ups. J. P. Metlay et al., "Measuring
symptomatic and functional recovery in patients with community-acquired pneumonia," J Gen Intern Med 12, no. 7
(1997).
16 Acumen, LLC
be more likely to result in a successful outcome in these regions. Although risk adjustment can
control for various beneficiary level characteristics including comorbidities, it is not possible to
perfectly control for the severity of a beneficiarys condition. Third, using diagnosis codes to
define condition cohorts may lead to an endogeneity problem: the lowest cost beneficiaries are
less likely to visit a physician, and therefore the methodology will assign fewer low cost
beneficiaries to a given condition cohort. If low cost beneficiaries are more common in certain
regions, the methodology will overstate costs in these regions for the condition cohorts. Finally,
this study does not have access to beneficiary medical records, and thus it is not possible to
independently verify diagnostic information. Previous research, however, has compared
diagnosis codes on Medicare claims against diagnosis codes included in medical records and
found a high positive predictive value (PPV).30,31,32
There are two possible methods of calculating beneficiary outcomes using the cohort-
based approach: including all claims within the observation window of the cohort and including
only claims that are related to each condition. Using all claims captures all health costs of a
beneficiary in a given cohort during the observation window. Including only claims that are
related to a condition requires the use of an episode grouper. Episode groupers use complex
algorithms to create clinically cohesive episodes of care that group treatment associated with a
single condition during the observation window. Because beneficiaries may have multiple
concurrent conditions, episode groupers must either divide claims among multiple episodes or
allocate claims to more than one episode.
This analysis includes all of a beneficiarys claims in the observation window. Using all
claims takes a holistic view by assuming that beneficiaries conditions affect all aspects of their
health, which may be particularly appropriate for chronic conditions. Though the all-claims
approach creates a single measure of a beneficiarys health, it produces false positives when
used with cohorts by including claims that are not necessarily related to each cohort. This
analysis attenuates the number of false positives by shortening the observation period for several
of the acute cohorts. Including claims unrelated to a given condition can be thought of as a
source of noise, but the large number of cases in each region ensures that it is unlikely that one
region will have more unrelated claims than another. The all-claims approach also double-
counts claims across cohorts when beneficiaries have concurrent conditions. Although double-
counting claims can overestimate total beneficiary outcomes, this analysis also creates an
30
Yuka Kiyota et al., "Accuracy of Medicare Claims-Based Diagnosis of Acute Myocardial Infarction: Estimating
Positive Predictive Value on the Basis of Review of Hospital Records," American Heart Journal 148(2004).
31
Wolfgang C Winkelmayer et al., "Identification of Individuals with CKD from Medicare Claims Data: A
Validation Study," American Journal of Kidney Diseases 46(2005).
32
Elena Birman-Deych et al., "Accuracy of ICD-9-CM Codes for Identifying Cardiovascular and Stroke Risk
Factors," Medical Care 43.
18 Acumen, LLC
3 MEASURING OF REGIONAL VARIATION IN HEALTH CARE
SPENDING, UTILIZATION AND QUALITY
To quantify the amount of geographic variation in the provision of health care services,
this study examines regional variation in health care expenditures, utilization, and quality of care.
Healthcare expenditures measure the actual costs incurred by each beneficiary, which includes
payments by insurers and beneficiaries. Geographic variation in expenditures, however, depends
on regional variation in both price and quantity. To more narrowly evaluate regional differences
in utilization patterns, this study measures regional variation in utilization as well. Utilization is
measured alternatively as a price-standardized spending levels (i.e., spending statistics that
control for regional differences in prices) or as a series count variables that measure differences
in the number of events per person (e.g., physician visits per capita) across regions. Areas where
beneficiaries incur high medical costs or use a large amount of services may not necessarily be
inefficient, however, if the increased use of medical services produces better health outcomes.
To determine whether or not this is the case, this study also uses a variety of quality measures to
assess the relationship between utilization of medical services and quality. Although this report
does not examine regional variation in quality measures directly, quality metrics for each HRR,
HSA and MSA are included in a supplemental appendix.
The following discussion presents the approach this study uses to measure regional
variation in these three outcome variables of interest. Sections 3.1, 3.2, and 3.3 describe how
this study defines health care spending, utilization, and quality outcomes, respectively. To
determine how the average values of these outcome measures varies across regions, the study
uses three definitions of a region. These definitions include Hospital Service Area (HSA),
Hospital Referral Region (HRR), and Metropolitan Statistical Area (MSA). Section 3.4
describes how these regions are constructed in detail.
The expenditure analysis examines regional variation in the costs of health care. With
health expenditures comprising 17.9 percent of the national Gross Domestic Product in 2010,
rising healthcare costs are a significant issue.33 Identifying areas that are high cost or low cost
may help policymakers implement initiatives to encourage low cost practice patterns. Regional
differences in healthcare spending per capita are due to two factors: differences in the price of
medical care and differences in the utilization of medical services. The expenditures analysis
investigates the overall change in spending but does not assess which of these two reasons is the
33
"National Health Expenditure Data," Centers for Medicare & Medicaid Services,
https://www.cms.gov/NationalHealthExpendData/downloads/tables.pdf.
The first step calculates the total raw cost during the beneficiary episodes. The total raw cost in
the Medicare analysis is calculated as the sum of the claim payment, coinsurance, and deductible
from the Inpatient, Outpatient, Hospice, Home Health, Skilled Nursing, Carrier, Durable Medical
Equipment, and Part D claim types. Expenditure calculations from the Inpatient claim type
remove the add-on Indirect Medical Education (IME) and Disproportionate Share (DSH)
payments. For Medicaid, total raw cost is calculated as the sum of the claim payment on FFS
claims found in the Inpatient, Prescription Drug, Long Term Care, and Other file types. In
addition, the Medicaid analysis excludes the costs from partial capitation claims, which appear
for beneficiaries in Primary Care Case Management (PCCM) programs.
Second, for all analyses, the expenditures variable adjusts raw costs for inflation between
years of analysis. The IOM Committee uses the average Consumer Price Index to adjust for
inflation, which takes into account the changing prices paid by consumers for all goods and
services.34 To inflate expenditures to costs in 2009, expenditures incurred in 2006 (for condition
cohorts that have a look-back period) are multiplied by 1.059; expenditures in 2007 are
multiplied by 1.032; expenditures in 2008 are multiplied by 0.995; and expenditures in 2010 are
multiplied by 0.986.
Finally, the analysis calculates per beneficiary, per month expenditures by dividing the
total costs for each beneficiary during his or her enrolled months by the number of months of
enrollment.35 A Medicare beneficiary is only considered enrolled during the months he or she is
enrolled in Parts A or B, and a Medicaid beneficiary is only considered enrolled during the
months he or she is FFS enrolled and with no benefit restrictions, as described in Section 2.2.5.
34
"Consumer Price Index - Chained Consumer Price Index," United States Department of Labor: Bureau of Labor
Statistics, http://bls.gov/data/.
35
In addition, the beneficiarys per month expenditures are weighted by the number of months of enrollment during
the risk adjustment analysis, which is described in Section 4.
20 Acumen, LLC
3.2 Measurement of Health Care Utilization
Examining utilization in addition to expenditures can reveal whether high cost regions are
expensive because of increased utilization of services rather than higher prices for these services.
This analysis measures regional variation of utilization in two ways:
Input price adjustment removes geographic variation in the price of inputs, such as labor costs,
and allows for an assessment of changing patterns of utilization overall. In addition to input-price
standardized cost, this study also measures the utilization of specific medical services as counts
of services. Input price adjustment gives a more comprehensive assessment of utilization than the
utilization counts by combining utilization from all sources and assigning appropriate relative
costs to each service (i.e., inpatient stays count for more than a physician visit does). Although
utilization counts are a less comprehensive measure than price-standardized cost, examining
specific service counts can better reveal regional variation for specific services that are most
likely to depend on patient or provider discretion. Section 3.2.1 defines the input price
standardized cost measure. Section 3.2.2 describes the counts of service utilization measures
included in this study.
Input price standardized cost assigns a standardized price for each service, so that the
price paid for each service is identical across all geographic regions. The analysis, in essence,
removes regional variation in Medicare and Medicaid payment rules to determine a base rate for
each service. For example, the Medicare analysis adjusts physician payments using the
Geographic Practice Cost Index (GPCI). The GPCI measures the relative cost of the inputs
physicians require to provide medical services against the national average input cost. The
Medicare analysis closely follows the standardization methodology developed in conjunction
with CMS as part of the Hospital Value-Based Purchasing (HVBP) program.36 Price
standardization occurs at the file type-year level, and prices are renormalized after
standardization so that the sum of the unadjusted and standardized costs are equal for a given file
type and a given year. Renormalization allows for the standardized costs to reflect the real prices
paid by CMS. For a technical description of Medicare input price standardization for all file
36
"Measure Methodology Reports: Medicare Spending Per Beneficiary (MSPB) Measure," QualityNet,
http://www.qualitynet.org/dcs/ContentServer?c=Page&pagename=QnetPublic%2FPage%2FQnetTier4&cid=122877
2057350.
While other subcontractors for the IOM geographic variation project calculate output
adjusted prices in addition to input adjusting costs, neither the Medicare nor Medicaid analysis
output-adjusts prices. Output price adjustment assigns each service the same cost regardless of
where it is performed by assigning services an average payment using Diagnosis-Related Groups
(DRGs) and CPTs. Medicares geographic adjustments are budget neutral and generally adjust a
standard national payment level for regional variation in the cost of providing each service.
Thus, because Medicares final prices are based on regional variation in input costs, using
separate output and input price adjustment is unnecessary. The final step renormalizes the
standardized price on each claim so that the total standardized payment for all claims equals the
total actual payment (excluding IME and DSH for inpatient)37 for each claim type in each year.
Medicaid price standardization also price-standardizes using input prices only. Although
Medicaid payments differ across the country because states have different payment systems, the
Medicaid analysis does not output-adjust prices for several reasons. First, about one third of
state Medicaid programs do not use DRGs to reimburse hospital providers, which precludes
output-standardizing prices using DRGs for those states.38 In addition, eight states do not
consistently utilize CPT codes but rely on local coding systems.39 The local coding systems do
not directly map to any nationally-used coding system, which prevents the analysis from
comparing specific services between geographic locations.
In addition to the overall input price adjusted expenditures, the input price adjusted
expenditures are also stratified by service categories. Stratifying by service category allows
analysis of different types of care, such as acute hospitalization or diagnostic services, to
determine if regional variation overall is similar to regional variation for specific types of
services. Appendix B.2 lists the service category specifications.
37
There are other special payments, such as those for Medicare Dependent Hospitals, Sole Community Hospitals,
Low Volume Hospitals, etc. Since these will all be part of total actual payment, but are not part of the standardized
price methodology, the renormalization will have the effect of uniformly raising the standardized price of inpatient
claims. Similar payments in other claim types, such as the add-on for physicians in shortage areas, will have similar
effects.
38
"Medi-Cal DRG Project: Frequently Asked Questions," California Department of Health Care Services,
http://www.dhcs.ca.gov/Documents/MCDrgProjectFAQs.pdf.
39
This analysis calculated the percentage of each states claims that are CPT claims and flagged eight states with
consistently low percentages of CPT use from 2008 through 2010. These states are New York, Connecticut, Idaho,
Maine, North Carolina, Ohio, Vermont, and Washington.
22 Acumen, LLC
3.2.2 Counts of Service Utilization
To analyze the geographic variability of utilization, this analysis also examines the
utilization counts of specific services. Price standardization can assess utilization across regions
by accounting for differences in input prices. Measuring utilization with count-based variables is
better able to measure regional variation in specific services that may be sensitive to supply- or
demand-side factors.
This methodology relies on codes from Medicare and Medicaid claims to calculate
measures such as the number of inpatient days with a surgical admission. For utilization
measures that rely on CPT or HCPCS codes, Medicaid data is underreported in states with
prevalent use of local coding systems, as discussed above. In addition, because of the structure of
Medicaid claims data, some related utilization measures are combined. For example, the
Medicaid methodology combines the number of inpatient surgical days and number of inpatient
medical days measures because medical and surgical admissions are defined by DRG, and
Medicaid claims in some states do not include DRGs. Appendix B.3 presents definitions of all
the utilization measures.
To analyze variation in the quality of care across geographic regions, the study utilizes a
broad range of well-established quality measures. The 18 measures chosen for this analysis are
all supported or developed by established institutions such as the National Committee for Quality
Assurance (NCQA), the Physician Consortium for Performance Improvement (PCPI), and the
Agency for Healthcare Research and Quality (AHRQ), in addition to other organizations. The
quality measures apply to specific conditions to determine how the quality of care for those
conditions varies. The methodology calculates separate quality measures for each cohort and for
the aggregate of beneficiaries in each region. For the aggregate cohort, this analysis employs
AHRQ composite quality indicators. Using these variables, this study can determine whether
areas with high utilization levels also have high quality medical care. Appendix B.4 lists the
condition-specific quality measures for the cohort analysis, and Appendix B.5 lists the composite
quality measures for the aggregate analysis.
Because certain states underutilize the DRG coding system in Medicaid, the composite
PSI measure is calculated differently for the Medicaid analysis than for the Medicare analysis.
Specifically, the individual PSI measures 8, 13, and 14 are left out of the composite regressions
and totals for Medicaid because they have an insufficient number of beneficiaries included in the
denominators of the measures. The remaining individual measures are reweighted. In addition,
for certain states, the composite PSI measure has an insufficient number of total beneficiaries in
the denominator to report meaningful rates. Thus, this analysis does not include in the regression
HRRs in states that do not utilize DRGs sufficiently. To determine the threshold of beneficiaries,
Figure 3.1 presents the ratio of the number of beneficiaries included in the composite PSI
measure denominator to the total number of beneficiaries by state. The percent of beneficiaries
included in the composite PSI measure denominator drops sharply after 18 percent to 5 percent.
Thus, this analysis considers states with fewer than 18 percent of their beneficiaries included in
the composite PSI denominator to have inadequate measure denominators to reliably represent
the quality of care, measured by the composite PSI, in that state.
24 Acumen, LLC
3.4 Defining a Region
Each of the analyses described in this report are performed using three different
geographic region definitions. These include:
40
"Research Methods: Dartmouth Atlas of Health Care," Dartmouth Atlas of Health Care,
http://www.dartmouthatlas.org/tools/faq/researchmethods.aspx.
41
United States Census Bureau, "Metropolitan and Micropolitan Statistcal Areas Main,"
http://www.census.gov/population/metro/.
The remainder of this section provides a more detailed discussion of this studys risk
adjustment model. Section 4.1 contains a formal presentation of the risk adjustment model used.
Section 4.2 describes the set of beneficiary-level and market-level variables that are used in
different regression specifications.
To account for geographic variation in beneficiary and market level characteristics, the
methodology risk adjusts all the outcome measures described above. The outcome measures are
treated as the dependent variable, and regional variation in the dependent variables is analyzed
using the residuals from the risk adjustment analysis. The risk adjustment analysis relies on an
average residuals approach. To implement this approach, the following specification is estimated
for each year of analysis:
(4.1)
In equation (4.1), , represents the dependent variable (expenditures, quality, or utilization) for
beneficiary i, represents the coefficients estimating the relationship between the dependent and
the independent variables represented by , and represents the error term. The dependent
variable represents the outcomes described in Section 3. This study uses an ordinary least
squares (OLS) regression methodology at the beneficiary level (or the event level for the disease
cohort analysis) to estimate the coefficients in equation (4.1).
The outcome variable, Yi, is calculated on a monthly basis. For example, in the
expenditure analysis, Yi represents average spending per month. Calculating average per-month
costs removes the effect of increased expenditures or utilization at different points in the
treatment of a condition, such as at the beginning. Weighting the outcomes gives more influence
26 Acumen, LLC
to beneficiaries who are enrolled for a longer period, in essence counting beneficiaries for each
of their months of enrollment. Although all risk adjustment models use this weighting scheme,
the discussion below assumes all observations are equally weighted for expositional clarity.
After estimating using OLS, the average residual is computed for each region in each
year using the following formula:
(4.2)
In equation (4.2), represents the average residual for region R, represents the residual for
beneficiary (or event) i, and NR represents the number of beneficiaries (or events) in region R.
The residual for individual i, , is calculated as the difference between the observed (or actual)
value of the dependent variable, , minus the predicted value of the dependent variable,
where denotes the OLS estimates of the coefficients, . The sum is
calculated for all beneficiaries in region R.
These risk adjustment factors are added to average spending to create a risk adjusted
value of spending for each region in each year, , which can be expressed as follows:
(4.3)
In equation (4.3), Y is equal to the average predicted value of the dependent variable across all
regions.
This approach is analogous to running a two-stage model. In the first stage, one runs the
regression in equation (4.1). In the second stage, one uses the individual level residuals, e, from
the first stage as the dependent variables in the second stage. The second stage regresses these
residuals on a series of HRR dummy variables.42
42
According to the Frisch-Waugh-Lovell Theorem, the regional spending estimates are equivalent to the regional
spending estimates, , from a fixed effects regression, (i.e,. ) if the risk adjusters, X, are linearly
unrelated to the regional dummy variables.
(4.4)
In equation (4.4), represents the dependent variable (the average utilization residual) for
region R, represents the coefficients for the market level independent variables represented by
for the beneficiaries region R, and represents the error term. Equation (4.4) is estimated
at the HRR region level.
To test model sensitivity and the impact of beneficiary and market level independent
variables on each dependent variable, the studies separately risk adjust dependent variables for
ten unique clusters43 of independent variables, which can be found in Appendix C.1 for the
Medicare analysis and Appendix C.2 for the Medicaid analysis. Cluster 2, the baseline model,
includes the following independent variables: beneficiary age, sex, health status (HCCs),
income/pharmacy benefit, partial-year enrollment, the interaction between age and sex, a new
enrollee indicator, and an indicator for the year of analysis. The baseline model independent
variables are chosen to be available to all subcontractors and promote consistency across
analyses. The cluster analyses allow researchers flexibility to evaluate model sensitivity and
answer diverse research questions when interpreting results of the aggregate analysis. For
example, the choice of whether to compare the risk-adjusted dependent variables using the
baseline model versus cluster 5, which includes additional race and income variables, depends on
whether the impact of race and income on those dependent variables is of interest. In addition,
certain independent variables are only accessible to particular IOM subcontractors as a result of
differential data sources and other factors, so various clusters are constructed with a common set
of variables to facilitate comparisons across subcontractor analyses while other clusters isolate
variables specific to selected subcontractors. For example, cluster 9 includes all independent
variables common to each subcontractor while cluster 10 includes independent variables unique
to the Medicare and Medicaid analysis. The results in this report use cluster 10 for the Medicare
analysis and cluster 5 for the Medicaid analysis.
Only the quality measures that are outcome measures are risk-adjusted. Outcome
measures assess the ultimate results of health care and must be risk adjusted to take into account
beneficiary factors. For example, the COPD admission rate should be adjusted for beneficiary
health status to avoid penalizing regions with beneficiaries that are sicker. Process measures
assess procedures that should be performed for all beneficiaries with a given condition, and thus,
they should not be adjusted for beneficiary factors. For example, all beneficiaries with diabetes
should have a screening for diabetic retinal disease, regardless of their gender or health status.
43
The term cluster in this report refers to regression specifications.
28 Acumen, LLC
The classifications of each quality measure can be found in Appendix B.4 and B.5. All quality
measures that are risk adjusted, other than the PQI measure for the aggregate cohort, use the
baseline model applied to a logistic regression rather than an OLS regression because the
outcomes are binary. The reported risk adjusted values for each episode are the observed
measure divided by the expected measure multiplied by the measure national mean across all
episodes. The HRR level mean for each quality measure is the average of these risk-adjusted
quality outcomes for each measure. The PQI composite measure, on the other hand, uses OLS
because the outcome is a count of events during the course of the observation window, while the
IQI and PSI composite measures use logistic regression in the same manner as the cohort
specific quality measures. The composite quality measures are not weighted to account for
precision due to the number of events in a region (as the AHRQ software does), so that the
analysis can observe all variation in quality.
This analysis risk adjusts spending, utilization, and quality across regions to control for
differences due to variation in patient case mix and in the price of medical care. Risk adjustment
controls for factors that can influence spending but are outside the providers control. The
independent variables used for risk adjustment were chosen in consultation with IOM and were
designed to be as consistent as possible across the Medicare and Medicaid analyses as well as the
commercial analyses performed by other subcontractors to IOM. To account for patient case
mix, the methodology employs a set of beneficiary-level characteristics which are described in
Section 4.2.1. In addition, some factors in the market can also affect spending and quality in a
region, such as the percent of people uninsured or the supply of physicians. Section 4.2.2
describes the set of market-level characteristics that this analysis uses for a separate set of
regression specifications.
Age
Sex44
Age-sex interaction
Race and ethnicity
Income
44
The prostate cancer cohort is not risk adjusted for beneficiary sex.
Risk adjusting for health status allows the methodology to account for the change in cost
associated with beneficiaries varying levels of health. The analysis utilizes CMS 2008
definition of Hierarchical Condition Categories (HCCs) as an indicator of health status. CMS
uses HCCs within its risk adjustment model used to determine capitation payments to Medicare
Advantage plans.45 HCCs aggregate ICD-9 diagnosis codes into clinical categories to determine
beneficiaries health status and comorbidities. Like the CMS-HCC model, this analysis examines
beneficiaries claims for the 12-month period prior to their observation start date to identify their
health status. Areas with higher utilization levels will also have more HCCs, as beneficiaries in
those areas have diagnoses recorded more often. Thus, beneficiaries in high-use areas will appear
somewhat sicker, and risk adjustment may bias the regression estimates toward a better (i.e.,
lower cost, higher quality) result.46 The Medicare analysis does not include the HCC interaction
terms that use dual-enrollment status. Though dual-enrollment status may be predictive of health
45
Gregory C. Pope et al., "Evaluation of the CMS-HCC Risk Adjustment Model," RTI International and the
Centers for Medicare & Medicaid Services(March 2011),
https://www.cms.gov/MedicareAdvtgSpecRateStats/downloads/Evaluation_Risk_Adj_Model_2011.pdf.
46
Medicare Payment Advisory Commission, "Regional Variation in Medicare Service Use: Report to Congress,"
(January 2011), http://www.medpac.gov/documents/Jan11_RegionalVariation_report.pdf.
30 Acumen, LLC
status, it is not, in itself, a health status indicator. Thus, the Medicare analysis includes a dual-
enrollment indicator in certain clusters but does not tie it to the health status indicators.
Applying the HCC model to the Medicaid population has a number of benefits and
disadvantages. Using the same approach to measure health status across payers increases the
consistency of the analysis. CMS designed the HCCs, however, to capture spending variation in
the Medicare population. It is likely that major diseases that affect the Medicare population also
affect many Medicaid beneficiaries as well. The diagnostic codes that trigger HCCs for the
Medicare population, however, may not necessarily be the ones that drive variation in cost for
the Medicaid population, since Medicaid beneficiaries are more likely to be younger and also
disabled. The degree of overlap between the diseases that Medicare beneficiaries suffer from
compared to those that Medicaid beneficiaries suffer from is largely outside the scope of this
project.
Finally, to account for other factors that influence the price of medical care, the
commercial subcontractors to the IOM utilize beneficiary-level employer and insurer
characteristics. However, because the Medicare and Medicaid claims data do not have
information regarding employer characteristics, the risk adjustment approach does not include
employer information.
To account for differences in spending, utilization, and quality due to the differing prices
of care across regions, the risk adjustment employs a set of market-level variables. These market-
level characteristics include:
Hospital competition
Percent of population uninsured
Supply of medical services per capita
Malpractice environment risk
Physician composition per capita
Access to care
Payer mix
Medicaid penetration
Health professional mix per capita
Percent of beneficiaries with supplemental Medicare insurance
These variables are meant to capture the underlying causes of the variation in prices of care. The
market-level characteristics are listed in Appendix 0. The market-level variables are drawn from
the following data sources:
This analysis uses an adjustment factor to calculate the per-capita health professional mix
variable from the ARF. Though the ARF includes a population estimate for most counties for
each year, some counties do not have a population estimate for a given year. For counties with
no population estimate for a year of analysis, the methodology inflates the countys 2000 United
States Census population count (which can be found in the ARF) assuming a growth rate equal to
the average national population growth rate:
,
(4.5) . , ,
,
In equation (4.5), Adj. Popj,t represents the adjusted population for county j in year t. Popj, t
represents the population estimate for county j as found in the ARF for year t, and jPopj,t
indicates the sum of the ARF population estimates for all counties, or the national population
estimate in the ARF, for year t. For example, for a county with no population estimate in 2007,
the analysis would calculate the adjusted population for that county as that countys population
in 2000 multiplied by the national population in 2007 and divided by the national population in
2000. This adjustment avoids underestimating the population of counties that do not have a
population estimate for a given year of analysis. In addition, if a county does not have data for a
given year for any market-level variable, the methodology uses the most recent data available.
As a result, an areas market-level variable could be the same for multiple years of analysis. For
example, if a county has data for 2008 but not for 2009 for a given market-level variable, the
2009 analysis will assign that variable the same value as the 2008 analysis.
47
"Area Resource File (ARF)," US Department of Health and Human Services, Health Resources and Services
Administration, Bureau of Health Professions, http://arf.hrsa.gov.
48
"Medicare Physician Fee Schedule (MPFS)," Centers for Medicare & Medicaid Services,
https://www.cms.gov/apps/physician-fee-schedule/overview.aspx.
49
HealthLeaders-InterStudy, http://hl-isy.com/.
50
American Hospital Association, http://www.aha.org.
51
"The National Profile of Local Health Departments Study Series," National Association of County & City Health
Officials, http://www.naccho.org/topics/infrastructure/profile/index.cfm.
32 Acumen, LLC
5 GEOGRAPHIC VARIATION IN MEDICARE SPENDING AND
UTILIZATION
Using the methodology described in the previous three chapters, this section answers six
research questions posed at the start of these report.
1. How much geographic variation exists in per capita volume of healthcare services?
2. Are regions with high Medicare utilization levels likely to have high utilization levels
in subsequent years?
3. Is the variation in the volume of Medicare services greater within or across regions?
4. Do regions that provide a high volume of Medicare services when treating
beneficiaries for a given disease also provide a high volume of medical services when
treating all other diseases?
5. What types of services are the primary drivers of regional variation in the utilization
of Medicare services?
6. Are areas with high utilization levels more likely to have high quality care?
Although the empirical analyses define regions alternatively as HRRs, MSAs and HSAs and
defines many clusters of analysis, for brevity this section only discusses analysis of regional
variation at the HRR level using cluster 10. The following six sections answer each of these
research questions in turn. Appendix D contains information describing the relationship between
Medicare quality and spending.
The distribution of Medicare spending across beneficiary criteria parallels that of other
major payers. Table 5.1 presents per capita monthly spending levels before price-standardization
or risk adjustment by beneficiary criteria for the aggregate cohort in the 2007 through 2009
analysis period.52 The distribution is heavily right-skewed as the median cost ($310) is far below
the mean cost ($964). In addition, there exist a large number of beneficiaries that do not incur
any Medicare expenditures over the year. In fact, over ten percent of all beneficiaries have $0
spending per month. The large number of beneficiaries with $0 spending and the right-skewed
52
Expenditures are calculated for each episode as average monthly expenditures. Each beneficiary may have more
than one episode per period of analysis. All analyses use expenditures weighted by the number of months the
beneficiary is enrolled during the observation window. For simplicity, the remainder of this report refers to the
episode-month as the beneficiary.
The right-skewed nature of the Medicare per capita spending levels persists even after
removing variation due to regional differences in prices, patient demographics, and observed
beneficiary severity of illness. Table 5.2 displays per capita spending levels after price-
standardization and risk adjustment. Because this analysis applies an OLS regression, risk
adjustment standardizes the average monthly cost across beneficiaries types identified as
explanatory variables in the model; this modeling produces an average monthly risk-adjusted
cost that is equal for all beneficiary criteria (except death/non-death, as this methodology did not
risk-adjust for episode outcome as a beneficiary-level variable). For the aggregate cohort, the
average price-standardized monthly cost is $958, while the median is $729, showing that
spending levels are still heavily right-skewed, but not as much as the unadjusted figures. The
skewness of the distribution can also be seen by contrasting the most extreme values with the top
and bottom 10th percentile; the difference between the 10th percentile and the minimum is under
53
Amy Finkelstein and Robin McKnight, "What did Medicare do? The initial impact of Medicare on mortality and
out of pocket medical spending," Journal of Public Economics 92, no. 7 (2008).
54
Naihua Duan et al., "A Comparison of Alternative Models for the Demand of Medical Care," Journal of Business
& Economic Statistics 1, no. April 1983, http://www.jstor.org/stable/1391852.
34 Acumen, LLC
$17,000, while the difference between the 90th percentile and the maximum is over $1,620,000.
Similar results can be found for the AMI, CHD, diabetes, and stroke episode cost distributions,
which are presented in Appendix D.1.
Although the effect of death on episode costs is uncertain, this analysis concludes that
death episodes are among the most expensive Medicare episodes. Episodes ending in death are
truncated and thus Medicare may avoid paying for services that the beneficiary would have
otherwise incurred had they lived through the entire year; on the other hand, beneficiaries who
die during the year may experience an increase in the utilization of intensive, high-cost
treatments that may be performed in an attempt to prolong the beneficiarys life. On average,
however, death episodes are high-cost; although only 4.4 percent of beneficiary episodes end in
death each year, over half of these beneficiaries fall into the top decile of unadjusted Medicare
spending. Although controlling for patient case mix decreases the share of death episodes in the
top decile of overall episode costs, death episodes still cost much more on average ($2,336) than
non-death episodes ($923).
The variability in the aggregate cost distribution across beneficiary criteria, measured
using the standard deviation and 90-10 difference, is highest among beneficiaries who died
during their episodes andto a lesser extentamong dual-eligible beneficiaries. Whereas the
Although not shown, this study reveals that the average monthly cost and variability for
beneficiaries with acute conditions generally is higher than for beneficiaries who have chronic
conditions. The average monthly cost in the year after a beneficiary has an AMI or stroke are
$5,591 and $5,047, respectively, and the average monthly costs for beneficiaries with CHD and
diabetes are $1,960 and $1,632, respectively, as presented in Appendix D.1. The standard
deviations of AMI and stroke are $4,710 and $3,928, and the standard deviation of CHD and
diabetes are considerably lower, at $2,538 and $2,260. Though the acute cohorts have higher
averages and more variability, the chronic cohorts have more high-cost outliers. These high-cost
outliers are more likely to occur for chronic episodes because the number of chronic episodes is
almost 18 times the number of acute episodes. In addition, though chronic conditions that are
well-managed can be low-cost, chronic conditions that are not well-managed can result in
extremely high costs.
If the same HRRs remain low volume regions across years, this finding suggests that
there exist persistent differences in provider practice patterns, beneficiary treatment preferences
or other time-invariant factors. On the other hand, if the lowest volume regions in one year are
high-volume the next, then variation in utilization rates may be the result of random noise rather
than any time-invariant cultural or institutional characteristics. To determine whether or not
medical service utilization is constant over time, the following analysis measures the correlation
of HRR average expenditures and rankings over time.
Within the years 2007 through 2009, relative utilization levels across HRRs are stable
over time. Figure 5.1 provides a scatterplot of each HRRs rank by price-standardized risk
adjusted cost for the aggregate cohort in 2007 and 2008. The figure indicates a strong
persistence of HRR utilization levels across time. Tables 5.4 and 5.5 quantify the strength of this
relationship. Table 5.3 shows that the Pearson correlation coefficients are above 0.95 for all
year-to-year comparisons of mean price-standardized, risk adjusted costs. Although not shown,
36 Acumen, LLC
the Pearson correlations are above 0.92 when using the median cost. The Spearman rank
correlation coefficients presented in Table 5.4 are similarly high. For all mean year-to-year
comparisons of the aggregate cohort the Spearman rank correlation coefficients are above 0.94.
Using median HRR cost, the Spearman rank correlations are above 0.91. This evidence indicates
that relative medical service volume across HRRs is stable over time; random variation is
unlikely to be generating outlier HRR utilization levels.
Figure 5.1: Medicare Utilization Rank by HRR, 2007-2008
Table 5.4: Spearman Rank Correlation for Standardized Risk-Adjusted Medicare Costs
2007 2008 2009
2007 1.000 0.963 0.942
2008 0.963 1.000 0.965
2009 0.942 0.965 1.000
38 Acumen, LLC
Figure 5.3: Medicare Utilization Ranks by HRR 2007-2008 (Cholecystectomy Cohort)
Though the correlation of costs and ranks over time is high for some cohorts, this analysis
cannot determine whether the correlation is due to provider culture, beneficiary preferences, or
other factors. Further, for the chronic cohort, the high levels of correlation may be due in part to
the fact that the same beneficiaries may appear in multiple chronic cohorts. Beneficiaries with
chronic conditions will have chronic conditions for life. Thus, the beneficiaries who have a
chronic condition in one year are largely the same beneficiaries who had the condition the year
before, whereas for acute conditions the cohorts are largely different sets of beneficiaries each
year. Although HRR utilization rankings are stable across the time period of this study, 2007
through 2009, further study is needed to identify whether these patterns persist over a longer time
frame.
Although the relative rankings of average HRR medical service utilization are persistent
over time, there exists substantial variation in health care expenditures and utilization within
regions. Specifically, health care expenditures and utilization differ across regions on average,
but regional differences in average spending and utilization are significantly smaller than
variation in beneficiary utilization levels within a given region. Throughout this section,
variation in spending and utilization calculated using data for beneficiaries residing in a given
region is referred to as within-HRR variation whereas across-HRR variation refers to
dispersion in average HRR spending and utilization across the nation. Average HRR spending
and utilization refers to the average over all beneficiaries residing in an HRR. The finding that
within-HRR variation is significantly larger than across-HRR variation indicates that it may be
While variation in spending and utilization within an HRR is considerably higher than
variation across HRRs, average HRR spending does explain a large share of variation in average
HSA spending. To estimate the share of variation in spending at the HSA level attributable to
HRRs, this analysis performs an OLS regression of HSA average spending on HRR indicator
variables, weighted by beneficiary-months enrolled in each HSA.56 The R2 value from this
regression represents the share of variation at the HSA level explained by HRR spending. An R2
value close to 0.0 would indicate that variation in HSA-level average spending is not at all
attributable to the HRR in which it is contained. An R2 value close to 1.0 would indicate that the
average spending is identical for each HSA in an HRR.
According to this methodology, more than three quarters of the variation in HSA
spending and utilization is explained by HRR spending. Table 5.6 presents the R2 value from the
regressions for expenditures and for utilization for the aggregate 2007 through 2009 analysis.
Seventy seven percent of the variation in HSA utilization is explained by the variation in HRR
utilization; 81 percent of the variation in HSA expenditures is explained by the variation in HRR
expenditures. Thus, the variation that remains in HSA-level average spending even after
controlling for HRR characteristics is less than 23 percent. This analysis does not take into
account, however, the variation in individual-level or provider-level expenditures and utilization
55
The analysis is limited to HRRs that contain 3 or more HSAs. In total, 285 out of 306 HRRs contain at least 3
HSAs.
56
In total, 285 out of 306 HRRs contain at least 3 HSAs.
40 Acumen, LLC
within an HSA. Thus, the high R2 on the regressions cannot be interpreted to indicate that HRR-
level spending adequately captures variation at the beneficiary or provider level.
Expenditures Utilization57
To further measure within- and across-HRR variation, this analysis employs two
measures of dispersion: standard deviation and the 90-10 difference. To determine within-HRR
variation, this analysis calculates the standard deviation and 90-10 difference using beneficiary
observations for beneficiaries residing in a given HRR. The standard deviations and 90-10
differences for each HRR are then averaged over all HRRs to generate two measures of within-
HRR dispersion. To determine across-HRR variation, this analysis calculates average spending
and utilization within each region using beneficiary observations for beneficiaries residing in that
region and measures the standard deviation and 90-10 difference of these HRR averages. Table
5.7 presents several measures of within- and across-HRR variation for unadjusted, price-
standardized, and price-standardized risk-adjusted expenditures. Unadjusted measures of
dispersion are misleading because they do not adjust for important differences across regions,
including differences in Medicare payment levels and in beneficiary characteristics. Risk
adjustment reduces variation due to beneficiary characteristics, resulting in lower within- and
across-region variation for the risk-adjusted costs.
Within-HRR variation in spending and utilization is an order of magnitude larger than the
across-HRR variation. The average standard deviation within HRRs of price-standardized risk-
adjusted expenditures is $1,621, whereas the standard deviation of average costs across HRRs is
$84. Similarly, the average 90-10 difference within HRRs is $2,094, while the 90-10 difference
across HRRs of the average HRR cost is $208. The weighted averages of the standard deviation
and 90-10 differences (which weight HRRs based on the count of beneficiary-months contained
in the HRR rather than treating each HRR equally) are even higher than the unweighted
averages. Within-region variation is also higher than across-region variation for unadjusted and
price-standardized expenditures.
The qualitative relationship between within-HRR and across-HRR variation also holds
for all condition cohorts; in particular, the AMI, stroke, CHD, and diabetes cohorts have
considerably higher within-HRR variation than across-HRR variation. Appendix D.4 presents
57
As in the remainder of this report, Medicare utilization is risk-adjusted using Cluster 10. Performing the
regression using Cluster 2 has similar results.
Table 5.7: Dispersion of Medicare Service Utilization Within and Across Regions
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $1,903 $1,900 $1,621
Standard Deviation of HRR Means $129 $135 $84
Average of HRR 90-10 Differences $2,479 $2,498 $2,094
90-10 Difference of HRR Means $304 $322 $208
To reconcile the observations that (i) there exists significant stability in HRR rank across
years and (ii) within-HRR variation is an order of magnitude larger than the across-HRR
variation, one can examine whether the typical or outlier beneficiaries are causing the
persistence in HRR utilization levels across time. Figures 5.7 and 5.8 present the rank of the
HRRs by mean price-standardized risk adjusted cost versus the rank by median and by the 90th
percentile cost, respectively, for the aggregate cohort for three years of analysis. While there is a
moderate level of correlation between an HRRs mean and median utilization levels, and mean
and median rank is somewhat high for the highest- and lowest-ranked HRRs, the relationship is
weak in general. The relationship between the mean and the 90th percentile rank, however, is
strong for all HRRs. These figures illustrate that stability of average HRR utilization levels
across years is due to the persistence in the cost of treating the highest-cost beneficiaries rather
than the cost of treating beneficiaries at the median.
42 Acumen, LLC
Figure 5.4: Medicare Utilization Levels by HRR, Mean vs. Median
Figure 5.5: Medicare Utilization Levels by HRR, Mean vs. 90th Percentile
To determine whether regions that have high utilization levels when treating beneficiaries
with a certain condition also have high utilization levels for treating beneficiaries with other
Overall, the highest unweighted Pearson correlation across HRRs of the average monthly
price-standardized risk-adjusted cost is between the COPD and CHF cohorts, at 0.947; the lowest
correlation is between the prostate cancer and cataract cohorts, at 0.230. Table 5.8 shows the
Pearson correlation coefficients across HRRs of the average monthly price-standardized residual
costs for the aggregate, AMI, stroke, CHD, and diabetes cohorts. The correlations between CHD,
diabetes, and the aggregate cohort are higher than the correlations between the AMI and stroke
cohorts. Table 5.9 displays the Spearman rank correlations for the same cohorts across HRRs,
ranked by the average monthly price-standardized residual cost, which show similar results. The
full comparison of Pearson correlations between all cohorts is presented in Appendix D.6.
The results indicate a higher correlation between beneficiaries with different chronic
conditions compared to beneficiaries with different acute conditions. The highest Pearson
correlation between chronic conditions is 0.947, between the COPD and CHF cohorts, while the
lowest correlation is 0.766, between the CHF and arthritis cohorts. Medicare beneficiaries often
have multiple chronic conditions, and the costs associated with each chronic illness will be
44 Acumen, LLC
counted toward all chronic episodes that are triggered. Beneficiaries with chronic conditions also
have them for multiple years, so the same beneficiaries are often included across years. In
addition, a single provider may care for a beneficiarys multiple chronic conditions, or a
beneficiary may see the same type of provider for his or her chronic conditions, such as a general
practitioner.
The HRR-level correlation of utilization levels for beneficiaries with acute conditions,
however, is typically much lower than the HRR-level correlation for chronic condition cohorts.
The lowest acute-condition correlation is 0.331, between the cataract and pneumonia cohorts.
Certain acute conditions do have higher correlations between HRRs; for instance the HRR-level
correlation between the AMI and pneumonia cohorts is 0.843. In general, the highest
correlations occur between conditions that are cared for by similar providers; for example,
beneficiaries with AMI and CHF are often treated by emergency room inpatient or outpatient
physicians and show a high correlation of 0.833. Although AMI and CHD are both diseases of
the heart, and the correlation between AMI and CHD is lower, at 0.742. This lower correlation
for these two heart-related diseases may occur because AMI and CHF are treated in the inpatient
setting whereas CHD is typically treated in the outpatient setting. This trend suggests that the
persistence in region spending is due, in part, to the high utilization levels of beneficiaries with
particular diseases who are cared for by the same types of providers.
Acute care
Post-acute care
Prescription drugs
Diagnostic
Acute care includes care provided in an inpatient setting, excluding inpatient psychiatric and
rehabilitation facilities; the post-acute care category is mostly home health, skilled nursing and
hospice care; and prescription drugs include drugs purchased under both the Medicare Part B and
Part D programs. The diagnostic service category includes physician visits to evaluate the
patient condition as well as medical test and imaging procedures. The procedures and
emergency department/ambulance categories are fairly self-explanatory and the other category
contains any claim not included in the first six groups. Appendix B.2 provides further details
about these groupings, which impose a hierarchy to create mutually exclusive classifications.
To create risk adjusted costs for each service category, this study re-estimates the OLS
risk model separately for each service category. The service-specific risk adjustment model uses
the same explanatory variables as the cross-service model. The risk adjusted costs for each
service category obtained from the risk model are weighted by the number of beneficiary-months
enrolled in the HRR. These weights allow HRRs with a greater number of observations to have
a larger impact on the results.
The post-acute and acute care risk adjusted costs have the strongest relationship with the
total average monthly risk adjusted costs. Figure 5.9 presents a series of charts of the average
monthly price-standardized cost residuals for each service category for the aggregate cohort. The
horizontal axis represents the HRRs sorted by their total average residual cost, with HRRs
furthest left being the lowest cost and HRRs furthest right being the highest cost. The vertical
axis shows the average monthly cost residual for the service category. If service utilization is
positively correlated with overall utilization, then it is expected that the graph will show a near-
linear relationship with a positive slope.
The results indicate that the utilization of post-acute care services is main driver of HRR-
level variation in utilization levels. Post-acute residual costs are closely related to total average
monthly residual costs; HRRs that have the lowest total cost residuals (on the left) also have the
lowest post-acute cost residuals, and HRRs that have the highest total cost residuals (on the right)
also have the highest post-acute cost residuals. The post-acute care service category also makes
up the largest share of the total cost residual, followed by the acute care service category. The
other categories make up small portions of the total cost residual and show little relationship with
total residual cost. Regional variation in the utilization of acute services also contributes to
46 Acumen, LLC
HRR-level variation in utilization levels, but regional variation in diagnostic, prescription drug,
procedure, ER/ambulance, and other monthly residual costs have little effect on HRR utilization
rankings.
3. Hospice care
Table B.2.1 in Appendix B defines these categories in more detail. This study estimates the OLS
risk adjustment model for each post-acute care category separately using the same set of
beneficiary-level covariates as the full model. As in the service category regressions, the risk
adjusted costs for each post-acute care type are weighted by the number of beneficiary-months
enrolled in the HRR.
Home health care has the largest geographic variation in spending of all types of post-
acute care. Table 5.10 presents the average monthly cost residual for each care category by
quintile. In Table 5.10, the HRRs are ranked by total utilization and categorized into quintiles
such that quintile 1 includes the 61 HRRs with the lowest total price-standardized, risk adjusted
spending, and each quintiles average residual is calculated for each post-acute care category.
Figure 5.10 graphically represents these results. Although skilled nursing care makes up the
largest share of post-acute spending, it has relatively low variability. Monthly home health care
costs, on average, less than half of monthly skilled nursing facility care, but the variability in
home health spending is much greater. To quantify the variability in each care setting, the final
column in Table 5.10 shows the range of the average monthly cost residual at the highest-cost
quintile (quintile 5) to the lowest-cost quintile (quintile 1). The high variability in home health
spending can also be seen graphically in Figure 5.10 as the difference between the absolute
values of the average monthly residual at the lowest-cost quintile and the average monthly
residual at the highest-cost quintile. Unlike other post-acute care settings, to be eligible for the
home health benefit, beneficiaries are not required to be hospitalized beforehand. Beneficiaries
become eligible for home health if they are confined to their home and require skilled nursing
care or physical or speech therapy on a part-time basis. Beneficiaries in home health do not pay
any cost-sharing and can receive an unlimited number of episodes of care as long as they are
HRRs with high utilization levels in one service category do not necessarily have high
utilization levels in other service categories. Although the correlation between the seven types of
58
Medicare Payment Advisory Commission, "Report to the Congress: Medicare Payment Policy," (March 2012),
http://www.medpac.gov/chapters/Mar12_Ch08.pdf.
50 Acumen, LLC
service categories is positive, it is fairly low. In fact, correlations between service category costs
are below 0.40. Table 5.9 presents the Pearson correlation of the average price-adjusted residuals
between each service category for the aggregate cohort in 2007. Table 5.9 reports the correlation
between each category and all other costs (the Remaining Costs category). The Remaining
Costs category is the total costs minus the costs of the corresponding service category for which
the correlation is reported in the cell. For example, the correlation between acute care costs and
remaining costs is defined as the correlation between acute care costs and total costs net of
acute care costs. If the Remaining Costs category had been reported as simply the total cost, its
correlation with the categories that are high-cost and make up a large share of the total cost
would appear artificially high. For the aggregate cohort, the other cost and prescription drug cost
categories tend to have lower correlation with other categories; the acute care and ER/ambulance
categories tend to have higher correlations. Appendix D.6 presents the Pearson correlations for
AMI, stroke, CHD, and diabetes. The chronic cohorts tend to have higher correlations overall
than the acute cohorts. The chronic cohorts also have patterns of correlation similar to the
aggregate cohort, a trend which is related to the high correlation between chronic cohorts and the
aggregate cohort described above.
Table 5.11: Medicare Service Category Utilization across HRRs, Pearson Correlation
(2007)
Prescription Drugs
Remaining Costs
Post-Acute Care
ER/Ambulance
Acute Care
Procedures
Diagnostic
Other
Remaining Costs . 0.28 0.06 0.21 0.24 0.12 0.31 0.08
Acute Care 0.28 1.00 0.03 0.13 0.27 0.04 0.29 0.05
Prescription Drugs 0.06 0.03 1.00 0.19 -0.01 0.16 0.02 0.05
Diagnostic 0.21 0.13 0.19 1.00 0.01 0.38 0.12 0.10
Post-Acute Care 0.24 0.27 -0.01 0.01 1.00 0.01 0.19 0.03
Procedures 0.12 0.04 0.16 0.38 0.01 1.00 0.05 0.08
ER/Ambulance 0.31 0.29 0.02 0.12 0.19 0.05 1.00 0.04
Other 0.08 0.05 0.05 0.10 0.03 0.08 0.04 1.00
1. How much geographic variation exists in per capita volume of healthcare services?
2. Are regions with high Medicaid utilization levels likely to have high Medicaid
utilization rates in subsequent years?
3. Is the variation in the volume of Medicaid services greater within or across regions?
4. Do regions that provide a high volume of Medicaid services when treating
beneficiaries for a given disease also provide a high volume of Medicaid services
when treating all other diseases?
5. What types of services are the primary drivers of regional variation in the utilization
of Medicaid services?
6. Are areas with high utilization levels more likely to have high quality care?
The first six questions are answered in the body of this text; Appendix E answers the seventh
research question regarding the relationship between quality and cost. After creating a detailed
picture of regional variation in Medicaid spending, utilization and quality, this section then
compares these results against the findings from the Medicare analysis. Specifically:
7. Are regions with high utilization levels in Medicare likely to have high utilization
levels in Medicaid?
The HRR-level results presented in this chapter rely on the methodology described in Sections 2
through 4 above.59 The following sections answer each of the above research questions in turn.
Beneficiaries that are female, Asian, or Hispanic are likely to have lower spending levels
than the average beneficiary. Table 6.1 presents per capita monthly spending levels before price-
standardization or risk adjustment by beneficiary criteria for the aggregate cohort during the
2007 through 2009 analysis period. The average cost per month for beneficiaries is $1,096,
implying that Medicaid beneficiaries cost $13,152 per year on average. Medicaid spending for
female beneficiaries, however, is about 60 percent of the spending levels for males. This result is
59
To be consistent with the Medicare analysis, cluster 5 is presented for the Medicaid analysis because it includes a
similar set of explanatory variables as the Medicare cluster 10.
52 Acumen, LLC
due in part to the fact that the average female enrollee is healthier than the average Medicaid
enrollee. For instance, state Medicaid programs are mandated to cover all pregnant women with
incomes up to 133 percent of the federal poverty line (FPL) and pregnant women represent a
generally healthier population than other Medicaid-eligible beneficiaries (e.g., the disabled). The
table below shows that women have 13.3 million episodes in the Medicaid cohort compared to
only 5.4 million episodes for men. In addition, Asian and Hispanic beneficiaries tend to have
lower average monthly expenditures than other racial groups. Similar to the Medicare cost
distribution, the overall Medicaid cost distribution is variable and right-skewed. As shown in the
table below, the standard deviation of monthly expenditures is $3,057. The spending
distributions right-skewed nature is demonstrated by the fact that the median cost for
beneficiaries in the aggregate cohort ($309) is far below the mean cost ($1,096). These patterns
holds for all beneficiary subgroups defined in the table below.
The Medicaid cost distribution remains highly variable and right-skewed even after price
standardization and risk adjustment. After removing variation due to regional differences in
prices, patient demographics, and observed beneficiary severity of illness, this skewness
becomes less pronounced (see Table 6.2). The average price-standardized monthly cost is
$1,094, while the median is $733. The long right tail of the Medicaid cost distribution is also
evident from the comparison of the difference between the 90th percentile cost and the maximum
cost and the difference between the 10th percentile and minimum cost; the largest difference for
any beneficiary category between the 90th percentile and the maximum is $1,805,090, while the
largest difference between the 10th percentile and the minimum is $21,780 (both for females).
Because this analysis applies an OLS regression including explanatory variables that identify
Similar to the aggregate cohort cost distribution, the cost distributions for the AMI,
stroke, CHD, and diabetes condition cohorts are right skewed. Further, average monthly cost
and variability for beneficiaries with acute conditions generally is higher than for beneficiaries
who have chronic conditions. Appendix E.1 displays the results for these four condition cohorts.
The average monthly cost in the year after a beneficiary has an AMI or stroke is $5,344 or
$4,737, respectively, and the average monthly cost for beneficiaries with CHD or diabetes is
$2,873 or $2,008, respectively. The standard deviations of AMI and stroke are $19,537 and
$15,701, and the standard deviation of CHD and diabetes are lower, at $12,236 and $8,674.
Medicaid utilization is fairly stable over time, but year-to-year correlations are much
weaker for Medicaid than for Medicare. Figure 6.1 provides a scatterplot of HRR rank by price-
standardized risk adjusted cost for the Medicaid aggregate cohort in 2007 and 2008, and Table
6.3 and Table 6.4 quantify the strength of this relationship. For all mean price-standardized, risk
adjusted cost year-to-year comparisons, the Pearson correlation coefficients range from 0.74 to
0.90 and range from 0.46 to 0.82 when using the median cost, which is not shown. For all mean
year-to-year comparisons, the Spearman rank correlation coefficients, presented in Table 6.4, are
between 0.74 and 0.91. Though not shown, the Spearman rank correlation coefficients using
median HRR cost range from 0.53 to 0.80. In contrast, the Pearson and Spearman correlation
coefficients for Medicare HRR rankings across all year combinations are all between 0.91 and
60
For some of the condition cohorts, the average monthly cost that is presented in Appendix E.1 is not equal across
beneficiary types. The beneficiaries in HRRs with fewer than 12 episodes for a given condition cohort are excluded
from the averages presented in these figures.
54 Acumen, LLC
0.98 when using either mean or median price-standardized, risk-adjusted HRR costs. This
evidence indicates that relative medical service volume across HRRs is far less stable over time
in the Medicaid program then in the Medicare program. Further analysis is needed to identify
whether this pattern of positive correlations and weaker Medicaid stability persists over a longer
time frame.
Table 6.4: Spearman Rank Correlation for Standardized Risk-Adjusted Medicaid Costs
2007 2008 2009
2007 1.000 0.905 0.738
2008 0.905 1.000 0.789
2009 0.738 0.789 1.000
Although regional Medicaid utilization levels are fairly stable over time, they are less
stable than Medicare utilization rates. There are a number of reasons for this finding. First, the
Medicaid data have smaller sample sizes in each HRR which may lead to increased volatility due
to outliers. Second, the Medicaid eligibility rules vary across States and are more likely to
change over time. For instance, eligibility levels for pregnant women currently range from the
Similar to the Medicare results, however, year-to-year stability in HRR rank is driven
primarily by the spending on the highest cost Medicaid beneficiaries rather than the typical
Medicaid beneficiaries. Figure 6.5 and Figure 6.6 plot HRR rank by mean price-standardized risk
adjusted cost versus the rank by median and by the 90th percentile cost, respectively, for the
aggregate cohort over the 2007 to 2009 period of analysis. These figures show that there is little
correlation between an HRRs mean and median utilization levels. The relationship between the
mean and the 90th percentile rank, however, is strong and positive, much like the Medicare
cohort. These figures illustrate that the stability of average HRR utilization levels across years is
due to the persistence in the cost of treating the highest-cost beneficiaries rather than the cost of
treating beneficiaries at the median of the cost distribution.
61
"Medicaids Role for Women Across the Lifespan: Current Issues and the Impact of the Affordable Care Act,"
The Henry J. Kaiser Family Foundation, http://www.kff.org/womenshealth/upload/7213-03.pdf.
62
"The Role of Section 1115 Waivers in Medicaid and CHIP: Looking Back and Looking Forward," The Henry J.
Kaiser Family Foundation, http://www.kff.org/medicaid/upload/7874.pdf.
56 Acumen, LLC
Figure 6.2: Medicaid Utilization Levels by HRR, Mean vs. Median
Figure 6.3: Medicaid Utilization Levels by HRR, Mean vs. 90th Percentile
Turning the focus to beneficiary subpopulations, HRR utilization levels for specific
beneficiary condition cohorts are positively correlated, butsimilar to the Medicare study
resultscorrelations are higher for beneficiaries with chronic conditions than for beneficiaries
with acute conditions. Figure 6.7 presents relative HRR rankings based on average monthly
58 Acumen, LLC
Figure 6.5: Medicaid Utilization Ranks by HRR, 2007-2008 (Cholecystectomy Cohort)
The highest-cost HSA within an HRR on average spends more than three times more per
Medicaid beneficiaries than the lowest cost HSA within that HRR. Table 6.5 displays the ratio of
highest- to lowest-spending HSAs within an HRR for HRRs that contain at least 3 HSAs.63 At
minimum, the lowest-spending or lowest-utilizing HSA uses only 1 percent more than the
highest-spending or highest-utilizing HSA within that HRR. At maximum, the highest-spending
(or highest utilizing) HSA in an HRR spends over 244 times more than the lowest-spending (or
lowest utilizing) HSA in that HRR. The variation in HSA-level spending within an HRR is
generally higher for Medicaid than for Medicaid.
Despite the large variation in HSA-level spending within certain HRRs, average HRR-
level spending on Medicaid beneficiaries explains more than half of the variation in HSA
spending. To reach this conclusion, this report estimates an OLS regression of HSA-level
average spending on HRR indicator variables weighted by beneficiary-months enrolled in each
63
The analysis is limited to HRRs that contain 3 or more HSAs. In total, 285 out of 306 HRRs contain at least 3
HSAs. The utilization range analysis also excludes an additional 6 HRRs with HSAs that have negative risk-adjusted
spending.
64
The maximum reported excludes one HRR with an HSA whose risk-adjusted spending was only $7, causing the
ratio of the highest to lowest spending to be inflated to a figure of more than 200.
Although HSA-level average spending levels have moderate levels of variability within
an HRR, Medicaid expenditures and utilization vary much more for beneficiaries within HRRs
and HSAs than across these regions. Table 6.8 presents several measures of within- and across-
HRR variation for unadjusted, price-standardized, and price-standardized risk-adjusted Medicaid
expenditures. The average standard deviation within HRRs of price-standardized risk-adjusted
expenditures is $2,446, whereas the standard deviation of average costs across HRRs is $334.
Similarly, the average 90-10 difference within HRRs is $2,229, while the 90-10 difference across
HRRs of the average HRR cost is $795. When examining unadjusted and price-standardized
expenditures, Table 6.8 shows that within-region variation also is higher than across-region
variation.67
Both within-HRR and across-HRR variation are higher for the Medicaid beneficiaries
than for Medicare beneficiaries. In particular, for the aggregate cohort, the average of HRR
standard deviations is over fifty percent higher for Medicaid than for Medicare and the standard
deviation of HRR average costs is over four times higher for Medicaid than for Medicare when
costs are price standardized and risk adjusted. Further, mean per-month Medicaid spending is
higher than mean Medicare spending.
65
In total, 285 out of 306 HRRs contain at least 3 HSAs.
66
As in the remainder of this report, Medicaid utilization is risk-adjusted using Cluster 5. Performing the regression
using Cluster 2 has similar results.
67
The weighted averages of the standard deviation and 90-10 differences (which weight HRRs based on the number
of beneficiary-months contained in the HRR rather than treating each HRR equally) are similar in magnitude to the
unweighted averages.
60 Acumen, LLC
diabetes cohorts have considerably higher within-HRR variation than across-HRR variation. For
example, the weighted average of HRR standard deviations for the AMI cohort is $16,973 while
the standard deviation of HRR mean costs is $1,632. Appendix E.4 presents within- and across-
HRR variation for each of these condition cohorts. Of the condition cohorts, the acute conditions
tend to have both larger within-region variation and larger across-region variation.
Table 6.7: Dispersion of Medicaid Service Utilization Within and Across Regions
Price-
Price-
Statistic Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $2,702 $2,706 $2,446
Standard Deviation of HRR Means $501 $497 $334
Average of HRR 90-10 Differences $2,772 $2,779 $2,229
90-10 Difference of HRR Means $1,126 $1,136 $795
Although HRRs that treat one beneficiary condition in a resource intensive manner are
likely to treat other beneficiary conditions cohorts in a resource intensive manner as well, this
finding is much more robust when comparing the treatment of chronic diseases rather than acute
diseases. To determine whether an HRR consistently uses above (or below) average level of
utilization levels to treat beneficiaries with specific conditions, the following analysis calculates
Pearson and Spearman rank correlations using the average monthly price-standardized residual
costs over all three years of analysis. Correlations between the chronic conditions range from
0.736 to 0.937; correlations between the chronic and acute conditions are generally smaller,
ranging from 0.522 to 0.744; and correlations between the acute conditions are generally the
smallest, ranging from 0.454 to 0.672. The highest correlation between chronic conditions is
between the low back pain and depression cohorts, at 0.937; the highest correlation between the
chronic and acute cohorts is between low back pain and cataract, at 0.744. The highest
correlation between acute conditions is between pneumonia and AMI, at 0.672. Table 6.9
presents selected correlation coefficients across HRRs for the aggregate, AMI, stroke, CHD, and
diabetes cohorts.68 Of the four condition cohorts, the highest correlation occurs between the two
chronic cohorts (CHD and diabetes); the lowest is between a the two acute cohorts (AMI and
stroke). Table 6.10 displays the Spearman rank correlation coefficients for these four conditions
and the aggregate cohort. The diabetes and CHD cohorts have the highest Spearman rank
correlation, at0.847, whereas the diabetes and AMI cohorts have the lowest, at 0.594.
68
Appendix E.6 presents a table of correlations between each cohort.
Aggregate 1.000
AMI 0.502 1.000
Stroke 0.515 0.568 1.000
CHD 0.651 0.693 0.607 1.000
Diabetes 0.794 0.576 0.611 0.831 1.000
Aggregate 1.000
AMI 0.420 1.000
Stroke 0.466 0.625 1.000
CHD 0.613 0.719 0.666 1.000
Diabetes 0.800 0.594 0.615 0.847 1.000
Condition pairs with similar patterns of geographic variation in relative utilization levels
for Medicare beneficiaries are also likely to exhibit similar patterns of relative resource intensity
for Medicaid beneficiaries. To illustrate this point, Figure 6.9 plots the Medicare correlation for
a given condition pair against the Medicaid correlation for that same condition pair. For example,
the correlation between AMI and stroke is 0.781 for Medicare and 0.568 for Medicaid; thus, the
point plotted at (0.781, 0.568) represents the relationship across Medicare and Medicaid of the
correlation between these two conditions. If the Medicare and Medicaid correlations were
identical, the points would lie on the 45 degree line shown in the graph. In general, the Medicaid
correlations are often lower than the Medicare correlations, but the graph shows a positive and
linear relationship, especially for condition pairs with higher correlations. The Pearson
correlation coefficient between the Medicare and Medicaid condition-pair correlations is 0.644,
indicating that there is a strong relationship between the condition correlations in Medicare and
Medicaid.
62 Acumen, LLC
Figure 6.6: Medicare vs. Medicaid Condition Correlations
HRRs with high resource intensity overall may not have an above average level of
resource utilization across all services. To determine whether or not high utilization HRRs can
be considered consistently high resource-intensity across all services, this analysis stratifies
price-standardized, risk adjusted expenditures by service category for each HRR. The following
analysis divides medical services into seven mutually exclusive, exhaustive categories. Section
6.6.1 defines each of these service categories and explains the methodology used to risk-adjust a
beneficiaries use of each of these medical services. Next, Section 6.6.2 examines which service
categories are the most important drivers of regional variation in overall resource intensity.
Finally, Section 6.6.3 examines whether geographic variation in overall utilization is correlated
with regional variation in the use of specific medical services.
This analysis decomposes Medicaid services into seven service categories. These
categories include:
Acute care
Post-acute care
Prescription drugs
Diagnostic
Procedures
Acute care includes care provided in an inpatient setting, excluding inpatient psychiatric and
rehabilitation facilities. The post-acute care category is mainly nursing home, home health,
hospice care, and care provided at rehabilitation and psychiatric facilities. Post-acute care in the
Medicaid setting is perhaps a misnomer because a large share of the cost in this service category
is from Medicaid residents of long-term care facilities, which may or may not be needed as a
result of care provided in the inpatient setting. For consistency across the Medicare/Medicaid
analyses, however, we have maintained the same naming conventions. Like in the Medicare
analysis, the diagnostic service category includes physician visits to evaluate the patient
condition as well as medical tests and imaging procedures. The groupings also impose a
hierarchy for the Medicaid analysis to create mutually exclusive classifications. Appendix B.2
provides the detailed definitions of these groupings, including the differences between the
Medicaid and Medicare specifications.
Risk adjustment for a beneficiarys use of medical services in each of these seven
categories follows the same approach as is used for the Medicare analysis. Specifically, this
study re-estimates the OLS risk model for each service category using the same explanatory
variables as the cross-service model. Because the aggregate and service-specific risk adjustment
models are linear and because the services are defined to be mutually exclusive and exhaustive,
the sum of the risk adjusted cost for each service category will be equal to the estimate risk
adjusted cost for overall spending.69
69
The risk adjusted costs are weighted by the number of beneficiary-months enrolled in the HRR, which allows
HRRs with greater numbers of beneficiary-months to have a greater influence on the results.
64 Acumen, LLC
total utilization. Of all the service categories, the post-acute care graph clearly shows the
strongest positive linear relationship with total utilization.
In addition to having the strongest relationship with total utilization, post-acute care has
the highest variation in expenditure levels across HRRs. To illustrate the variation in post-acute
care spending, an HRR in the 10th percentile of post-acute care spending spends $100 less than
average on post-acute care, whereas an HRR at the 90th percentile of post-acute spending spends
$157 more than average. Furthermore, the graph suggests that areas that are high spending in the
post-acute category are also high-spending overall.
Post-acute services are a main driver of regional variation in total Medicaid utilization
levels, but there does exist a weaker positive relationship between regional variation in the use of
specific types of medical services and overall service utilization in an HRR. For the highest cost
HRRs, the use of prescription drugs and medical services in the other category make up a large
share of an HRRs overall utilization levels. In general, compared to the Medicare results, the
Medicaid results show that a number of service categories affect the total cost residual, and the
service categories that drive high costs in HRRs are not the same service categories that drive
low costs in HRRs.
Figure 6.7: Medicaid Service Category Standardized Risk Adjusted Residual, by HRR
To change practice patterns for high cost HRRs, policymakers need to determine if above
average resource intensity is driven by the practice patterns of all providers or driven by a subset
of providers supplying specific medical services. To determine whether high cost areas have
consistently high utilization levels across all service categories, Table 6.11 shows the Pearson
correlation of the average price-adjusted residuals between each service category for the
aggregate cohort in 2007. The remaining costs category in Table 6.11 is the total costs in the
HRR minus the costs of the service category with which the correlation is reported. Examining
the correlation of each service category with the remaining costs service category prevents
artificially inflating the estimated correlations for service categories that make up a large share of
total cost.
HRRs that have high utilization levels do not necessarily have high utilization levels
across all service categories. The correlations between the seven service categories are always
positive and range from 0.005 to 0.311, indicating that there is some relationship between service
categories. The diagnostic and procedures categories tend to have the highest correlations with
other categories overall. Post-acute care has some of the lowest correlations with other
categories even though it makes up the largest share of the expected cost residual. The results
are similar for the 2008, 2009 and 2007-2009 specifications as well. Appendix E.5 presents the
Pearson correlations for the AMI, stroke, CHD, and diabetes cohorts, which show similar results.
Prescription Drugs
Remaining Costs
Post-Acute Care
ER/Ambulance
Acute Care
Procedures
Diagnostic
Other
Remaining Costs .
Acute Care 0.08 1.00
Prescription Drugs 0.05 0.01 1.00
Diagnostic 0.31 0.23 0.09 1.00
Post-Acute Care 0.02 0.01 0.02 0.01 1.00
Procedures 0.22 0.18 0.09 0.28 0.00 1.00
ER/Ambulance 0.07 0.05 0.02 0.26 0.01 0.12 1.00
Other 0.05 0.02 0.07 0.06 0.00 0.06 0.02 1.00
Overall, the relationships between service categories tend to be stronger for Medicare
than Medicaid. In addition, Medicaid post-acute care utilization has a lower correlation with
other services than does Medicare post-acute care utilization. This finding is likely due to the
fact that Medicare only pays for short-term skilled nursing home care following a hospital stay,
whereas Medicaid does not have this requirement. Although the acute care and ER/ambulance
categories have the highest correlations in Medicare, the diagnostic and procedures categories
have the highest correlations in Medicaid. Correlations across the service categories tend to be
higher for the chronic cohorts than for the acute cohorts in both Medicare and Medicaid.
Regions that are high-cost for Medicare beneficiaries are not necessarily high-cost for
Medicaid beneficiaries. Using the price-standardized risk-adjusted cost from 2007 through 2009,
Figure 6.11 presents the ranks of each HRR in Medicare on the horizontal axis and their ranks in
Medicaid on the vertical axis.70 To quantify the relationship between Medicare and Medicaid
utilization, Table 6.12 calculates the Pearson correlation and Spearman rank correlation
coefficients between the price-standardized, risk-adjusted expenditures in Medicare and
Medicaid for each HRR. Both correlation statistics produce slightly negative estimates; the
Pearson is -0.07, and the rank correlation is -0.06. The finding that correlation is near zero
indicates that regions with high expenditures in Medicare do not necessarily have high
expenditures in Medicaid.
70
The size of the bubble that represents each HRR is proportional to the number of Medicaid episodes in the
aggregate cohort in that HRR from 2007 through 2009.
68 Acumen, LLC
Figure 6.8: Medicare versus Medicaid Ranks
1. How much geographic variation exists in per capita volume of healthcare services?
2. Are regions with high utilization levels likely to have high utilization rates in
subsequent years?
3. Is the variation in the volume of medical services greater within or across regions?
4. Do regions that provide a high volume of medical services when treating beneficiaries
for a given disease also provide a high volume of medical services when treating all
other diseases?
5. What types of services are the primary drivers of regional variation in the utilization of
medical services?
6. Are areas with high utilization levels more likely to have high quality care?
7. Are regions with high utilization levels in Medicare likely to have high utilization
levels in Medicaid?
The following two sections present a summary of findings for the Medicare and Medicaid
investigations respectively.
This analysis reveals that even after adjusting for geographic price variation and
beneficiary health status, costs vary significantly across regions. Whereas the average Medicare
patient uses $958 worth of services each month, patients in Rochester, NY, use only $784 of
medical services. If Medicare could reduce utilization levels to those of the most efficient HRR
(i.e., Rochester, NY), it would reduce costs by $68 billion per year. Even if Medicare could
reduce average beneficiary utilization levels to those of the HRR at the 25th percentile (St.
Cloud, MN), it would save Medicare $24.5 billion per year.
Not only is there significant variation in the utilization of healthcare services in Medicare,
high utilization regions tend to remain so over time. The correlation of HRR-level utilization in
Medicare between 2007 and 2008 is 0.97. This strong year-to-year correlation, however, is not
driven by similar treatment patterns for the typical patient over time; rather the correlation is
70 Acumen, LLC
due in large part to persistence in the cost of treating the highest-cost beneficiaries across
regions. The correlation of average and 90th percentile utilization levels is 0.93 whereas the
correlation of average and median utilization levels is only 0.70.
Like Medicare, HRRs that have high intensity of resource use in one year are likely to
have high intensity of resource use in the following year. The year-to-year correlation of HRR-
level utilization for Medicaid (0.90) is comparable, although somewhat smaller than for
Medicare (0.97). Like Medicare, the variability of Medicaid expenditures and utilization is
significantly higher among beneficiaries within a given HRR than for the average beneficiary
across HRRs. The standard deviation of Medicaid spending for patients within a given HRR is
over 5 times larger than the standard deviation of Medicaid spending for the average beneficiary
across HRRs. Finally, like Medicare, a broad-based approach to achieving cost savings will fail
to take into account that high cost areas are not necessarily high cost for treating patients with all
diseases.
Although the geographic variation in Medicaid resource use exhibits similar features as
the geographic variation in Medicare use, there exists practically no statistical relationship
between an HRRs Medicare utilization levels and their Medicaid utilization levels. The
correlation between an HRRs Medicare and Medicaid utilization levels is -0.07. Table 7.1
provides a high-level comparison of key statistics results of this reports analysis of regional
variation in Medicare and Medicaid spending and utilization.
71
Positive numbers indicate higher spending is correlated with higher quality.
72 Acumen, LLC
REFERENCES
"2007 State Perspectives Medicaid Pharmacy Policies and Practices." National Association of
State Medicaid Directors.
"Acceptable Physician Specialty Types for Risk Adjustment Data Submission." Centers for
Medicare & Medicaid Services, http://www.csscoperations.com/internet/Cssc.nsf/files/Risk-
Adjust-Physn-Spec-Types031511_040811.pdf/$FIle/Risk-Adjust-Physn-Spec-
Types031511_040811.pdf.
"Area Resource File (ARF)." US Department of Health and Human Services, Health Resources
and Services Administration, Bureau of Health Professions, http://arf.hrsa.gov.
"Berenson-Eggers Type of Service (BETOS)." Centers for Medicare & Medicaid Services,
https://www.cms.gov/Medicare/Coding/HCPCSReleaseCodeSets/BETOS.html.
"CMS Manual System: Pub 100-20 One-Time Notification." Centers for Medicare & Medicaid
Services, http://www.cms.gov/Regulations-and-
Guidance/Guidance/Transmittals/downloads/R978OTN.pdf.
"Consumer Price Index - Chained Consumer Price Index." United States Department of Labor:
Bureau of Labor Statistics, http://bls.gov/data/.
"Consumer Price Index - Chained Consumer Price Index." United States Department of Labor:
Bureau of Labor Statistics, http://bls.gov/data/.
"IPPS Annual Proposed and Final Rules." Centers for Medicare & Medicaid Services,
http://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/AcuteInpatientPPS/.
"Medicaid - a Primer: Key Information on Our Nations Health Coverage Program for Low-
Income People." The Kaiser Commission on Medicaid and the Uninsured,
http://www.kff.org/medicaid/upload/7334-04.pdf.
"Medicaid Managed Care: Key Data, Trends, and Issues." Kaiser Commission on Medicaid and
the Uninsured, http://www.kff.org/medicaid/upload/8046-02.pdf.
"Medicaids Role for Women across the Lifespan: Current Issues and the Impact of the
Affordable Care Act." The Henry J. Kaiser Family Foundation,
http://www.kff.org/womenshealth/upload/7213-03.pdf.
"The Medicare Advantage Program." MedPAC Report to the Congress: Medicare Payment
Policy, http://www.medpac.gov/chapters/Mar09_Ch03.pdf.
"The Medicare Advantage Program: Status Report." MedPAC Report to the Congress: Medicare
Payment Policy, http://www.medpac.gov/chapters/Mar12_Ch12.pdf.
"Medicare Advantage Rates & Statistics: Risk Adjustment." Centers for Medicare & Medicaid
Services, https://www.cms.gov/Medicare/Health-
Plans/MedicareAdvtgSpecRateStats/Risk_adjustment.html.
"Medicare Claims Processing Manual: Chapter 26 - Completing and Processing Form CMS-
1500 Data Set. Section 10.8.2: Physician Specialty Codes." The Centers for Medicare &
Medicaid Services, http://www.cms.gov/Regulations-and-
Guidance/Guidance/Manuals/Downloads/clm104c26.pdf.
"Medicare Demonstration Projects & Evaluation Reports." Centers for Medicare & Medicaid
Services, http://www.cms.gov/Medicare/Demonstration-
Projects/DemoProjectsEvalRpts/index.html.
"Medicare Economic Index Web Table." Centers for Medicare & Medicaid Services,
http://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-Trends-and-
Reports/MedicareProgramRatesStats/downloads//mktbskt-economic-index.pdf.
"Medicare Managed Care Manual, Chapter 1: Section 1876 Cost Plans." Centers for Medicare &
Medicaid Services, http://www.cms.gov/Regulations-and-
Guidance/Guidance/Manuals/Downloads/mc86c01.pdf.
"Medicare Managed Care Manual, Chapter 8: Budget Neutral (BN) Risk Adjustment." Centers
for Medicare & Medicaid Services, http://www.cms.gov/Regulations-and-
Guidance/Guidance/Manuals/Downloads/mc86c08.pdf.
"Medicare Managed Care Manual, Chapter 8: CMS' Payments to Hospice Programs." Centers for
Medicare & Medicaid Services, http://www.cms.gov/Regulations-and-
Guidance/Guidance/Manuals/Downloads/mc86c08.pdf.
"Medicare Physician Fee Schedule (MPFS)." Centers for Medicare & Medicaid Services,
https://www.cms.gov/apps/physician-fee-schedule/overview.aspx.
74 Acumen, LLC
"Medicare: Medicare Advantage Fact Sheet." The Henry J. Kaiser Family Foundation,
http://www.kff.org/medicare/upload/2052-15.pdf.
"National Health Expenditure Data." Centers for Medicare & Medicaid Services,
https://www.cms.gov/NationalHealthExpendData/downloads/tables.pdf.
"The National Profile of Local Health Departments Study Series." National Association of
County & City Health Officials,
http://www.naccho.org/topics/infrastructure/profile/index.cfm.
"Quality Indicator User Guide: Inpatient Quality Indicators (IQI) Composite Measures, Version
4.4." The Agency for Healthcare Research and Quality,
http://qualityindicators.ahrq.gov/Downloads/Modules/IQI/V44/Composite_User_Technical_
Specification_IQI%20V4.4.pdf.
"Quality Indicator User Guide: Patient Safety Indicators (PSI) Composite Measures, Version
4.4." The Agency for Healthcare Research and Quality,
http://qualityindicators.ahrq.gov/Downloads/Modules/PSI/V44/Composite_User_Technical_
Specification_PSI%20V4.4.pdf.
"Quality Indicator User Guide: Prevention Quality Indicators (PQI) Composite Measures,
Version 4.4." The Agency for Healthcare Research and Quality,
http://qualityindicators.ahrq.gov/Downloads/Modules/PQI/V44/Composite_User_Technical
_Specification_PQI%20V4.4.pdf.
"Research Methods: Dartmouth Atlas of Health Care." Dartmouth Atlas of Health Care,
http://www.dartmouthatlas.org/tools/faq/researchmethods.aspx.
"The Role of Section 1115 Waivers in Medicaid and Chip: Looking Back and Looking
Forward." The Henry J. Kaiser Family Foundation,
http://www.kff.org/medicaid/upload/7874.pdf.
Anthony, Stephanie E., Christy Schroer, Sharon Silow-Carroll, and Jack A. Meyer. "Medicaid
Managed Care for Dual Eligibles: State Profiles." In, The Kaiser Commission on Medicaid
and the Uninsured 14, (2000).
http://www.kff.org/medicaid/loader.cfm?url=/commonspot/security/getfile.cfm&PageID=13
759.
Centers for Medicare & Medicaid Services. "Medicare Managed Care Manual, Chapter 1:
Introduction." http://www.cms.gov/Regulations-and-
Guidance/Guidance/Manuals/Downloads/mc86c01.pdf.
Duan, Naihua, Jr. Willard G. Manning, Carl N. Morris, and Joseph P. Newhouse. "A
Comparison of Alternative Models for the Demand of Medical Care." In, Journal of
Business & Economic Statistics 1, no. April 1983: 115-26.
http://www.jstor.org/stable/1391852.
Finkelstein, Amy, and Robin McKnight. "What Did Medicare Do? The Initial Impact of
Medicare on Mortality and out of Pocket Medical Spending." Journal of Public Economics
92, no. 7 (2008): 1644-68.
Franzini, Luisa, Osama I. Mikhail, and Jonathan S. Skinner. "McAllen and El Paso Revisited:
Medicare Variations Not Always Reflected in the under-Sixty-Five Population." Health
Affairs 29, no. doi: 10.1377/hlthaff.2010.0492 (2010): 2302-09.
Gawande, Atul. "The Cost Conundrum: What a Texas Town Can Teach Us About Health Care."
In, New Yorker (June 2009).
http://www.newyorker.com/reporting/2009/06/01/090601fa_fact_gawande.
Gold, H. T., and H. T. Do. "Evaluation of Three Algorithms to Identify Incident Breast Cancer in
Medicare Claims Data." In, Health Serv Res 42, no. 5 (2007): 2056-69.
http://www.ncbi.nlm.nih.gov/pubmed/17850533.
HealthLeaders-InterStudy. http://hl-isy.com/.
HEDIS 2010 Final NDC Lists. "Table CDC-A: Prescriptions to Identify Members with
Diabetes." National Committee for Quality Assurance,
http://www.ncqa.org/tabid/1091/Default.aspx.
Kiyota, Yuka, Sebastian Schneeweiss, Robert J Glynn, Carolyn Cannuscio, Jerry Avorn, and
Daniel H Solomon. "Accuracy of Medicare Claims-Based Diagnosis of Acute Myocardial
Infarction: Estimating Positive Predictive Value on the Basis of Review of Hospital
Records." American Heart Journal 148 (2004): 99-104.
Lakdawalla, Darius, Tomas Philipson, and Dana Coldman. "Addressing Geographic Variation
and Health Care Efficiency: Lessons for Medicare from Private Health Insurance." In, AEI
Health Policy Outlook 2, no. July (2010). http://www.aei.org/docLib/2010-7-No-2-g.pdf.
76 Acumen, LLC
Levinson, Daniel R., Inspector General. "Medicaid Managed Care Encounter Data: Collection
and Use." In, 12, no. OEI-04-07-00240 (2009). http://oig.hhs.gov/oei/reports/oei-07-06-
00540.pdf.
MaCurdy, T., J. Shafrin, D. Zheng, and F. Thomas. "Optimal Pay-for-Performance Scores: How
to Incentivize Physicians to Behave Efficiently."
http://www.cms.gov/reports/downloads/MaCurdy_Incentivize_Physicians_Optimal_P4P_Sc
ores_Feb_2011.pdf.
. "Regional Variation in Medicare Service Use: Report to Congress." In, (January 2011).
http://www.medpac.gov/documents/Jan11_RegionalVariation_report.pdf.
Merlis, Mark. "Medicare Advantage Payment Policy." National Health Policy Forum (2007).
Pope, Gregory C., John Kautter, Melvin J. Ingber, Sara Freeman, Rishi Sekar, and Cordon
Newhart. "Evaluation of the CMS-Hcc Risk Adjustment Model." In, RTI International and
the Centers for Medicare & Medicaid Services (March 2011).
https://www.cms.gov/MedicareAdvtgSpecRateStats/downloads/Evaluation_Risk_Adj_Mod
el_2011.pdf.
The Center for the Evaluative Clinical Services, and Dartmouth Medical School. The Dartmouth
Atlas of Health. Chicago: American Hospital Publishing, Inc., 1996.
United States Census Bureau. "Metropolitan and Micropolitan Statistcal Areas Main."
http://www.census.gov/population/metro/.
78 Acumen, LLC
Appendix A: COHORT DEFINITIONS
A.1 Clean Period Requirements
The table below classifies each condition as chronic or acute and lists the clean period
requirement. Section A.2 lists the source of the codes used to identify beneficiaries in each
cohort and describes the definition of the index date, the length of the observation period, and the
length of the clean period if there is a clean period requirement. Section A.3 lists all codes
relevant to each condition cohort definition.
Acute Myocardial
Acute 60 days
Infarction
The table below contains the algorithms for specifying the condition cohorts for the
Medicare study. The Medicaid study utilizes identical algorithms in most cases, however certain
CPT and HCPCS codes are not reported using uniform coding across states, and cohort
definitions utilizing these codes may underreport the number of beneficiary members of these
cohorts in the Medicaid analysis.
Code
Condition Medicare Study Cohort Algorithm
Source
Acute/ ICD-9 Dx Index date: At least 1 hospital inpatient admission with Dx code.
Ischemic Need 60 day clean period prior to index date.
Stroke Observation period: 12 months after index date.
Position: 1
Age: 18+
Diabetes ICD-9 Dx, Index date: 1 inpatient or 2 outpatient with Dx code OR 1 outpatient
NDCs and drugs
Observation period: 12 months after index date.
Position: 1
Age: 18+
Pneumonia ICD-9 Dx Index date: 1 inpatient or 2 outpatient Dx codes; 90 day clean period
prior to incident.
Observation period: 90 days after index date.
Position: 1
Age: 18+
Rheumatoid ICD-9 Dx, Index date: 1 inpatient or 2 outpatient with Dx code
Arthritis DRG Observation period: 12 months after index date
Position: 1-3
Age: 18+
Depression ICD-9 Dx Index date: 1 inpatient or 2 outpatient with Dx code
Observation period: 12 months.
Position: 1-3
Age: 18+
Congestive ICD-9 Dx Index date: Inpatient or 2 outpatient Dx code
Heart Observation period: 12 months
Failure Position: 1-3
Age: 18+
Acute ICD-9 Dx Index date: 1 inpatient with Dx code. Need 60 day clean period
Myocardial prior to index date.
Infarction Observation period: 12 months
Position: 1-3
Age: 18+
Coronary ICD-9 Dx Index date: 1 inpatient or 2 outpatient with Dx code
Heart Observation period: 12 months.
Disease Position: 1-3
Age: 18+
COPD ICD-9 Dx, Index date: 1 inpatient or 2 outpatient with Dx code or HCPCs
CPT Codes, Observation period: 12 months
& HCPCS Position: 1-3
Age: 18+
80 Acumen, LLC
Code
Condition Medicare Study Cohort Algorithm
Source
Cataract ICD-9 Index date: First occurrence of procedure code
Procedure, Observation period: 3 months after the first (or second) procedure
CPT, ICD-9 code. Thus, the observation period could range anywhere from 3 to 6
Dx months.
Position: 1-3
Age: 18+
Low Back ICD-9 Dx, Index date: 1 inpatient or 2 outpatient Dx codes
Pain ICD-9 Observation period: 12 months
Procedure Position: 1-3
codes Age: 18+
Chole- ICD-9 Index date: 1 inpatient procedure code
cystectomy Procedure, Observation period: 90 days prior to and after the index date.
CPT Codes, Position: 1-3
DRG Age: 18+
Breast ICD-9 Dx, Index date: 1 Medpar, Outpatient, or Carrier Claim Dx code and 1
Cancer ICD-9 procedure code from Medpar, Outpatient, or Carrier Claims within a
Procedure, window of 90 days before or after the first Dx (Diagnosis) date.
CPT Codes, Observation period: 12 months
& HCPCS Position: 1-3
Age: 18+
Prostate ICD-9 Dx, Index date: 2 Medpar, Outpatient, or Carrier Claim Dx code with
Cancer HCPCS, & different dates of service and 1 procedure code from Medpar,
CPT Codes Outpatient, or Carrier Claims within a window of 30 days before or
after the first Dx (Diagnosis) date.
Observation period: 12 months
Position: 1-3
Age: 18+
Female beneficiaries are excluded from this cohort.72
Lung ICD-9 Dx Index date: At least one Dx code on at least two claims with
Cancer different dates of services, at least one of which is primary, within
the index year anywhere on the Medpar.
Observation period: 12 months
Position: 1-3
Age: 18+
72
Out of 224,580 Medicare beneficiaries in the aggregate prostate cancer cohort, 10 female beneficiaries were
removed from the cohort due to this restriction. Out of 3,510 Medicaid beneficiaries in the prostate cancer cohort, 2
female beneficiaries were removed.
NDC Codes73
480.0 Pneumonia (Pn) due to adenovirus
480.1 Pn due to respiratory syncytial virus
480.2 Pn due to parainfluenza virus
480.3 Pn due to SARS-associated coronavirus
480.8 Pn due to other virus not elsewhere classified
480.9 Viral Pn unspecified
481 Pneumococcal Pn
482.0 Pneumonia due to klebsiella pneumoniae
482.1 Pn due to pseudomonas
482.2 Pn doe to haemophilus influenzae
Pneumonia
482.3x Pn due to streptococcus (excludes 481)
482.4x Pn due to staphylococcus
482.8x Pn due to other specified bacteria
482.9 Pn due to Bacterial pneumonia unspecified
483.0 Pn due to Mycoplasma pneumoniae
483.1 Pn due to chlamydia
483.8 Pn due to other specified organism
485 Bronchopneumonia organism unspecified
486 Pn organism unspecified
487.0 Influenza with Pn
73
HEDIS 2010 Final NDC Lists, "Table CDC-A: Prescriptions to Identify Members with Diabetes," National
Committee for Quality Assurance, http://www.ncqa.org/tabid/1091/Default.aspx.
82 Acumen, LLC
Condition Codes
296.20 Major depressive disorder single episode unspecified degree
296.21 Major depressive affective disorder single episode mild degree
296.22 Major depressive affective disorder single episode moderate
296.23 Major depressive affective disorder single episode severe degree without
psychotic behavior
296.24 Major depressive affective disorder single episode severe degree specified as
with psychotic behavior
296.25 Major depressive affective disorder single episode in partial or unspecified
remission
296.26 Major depressive affective disorder single episode in full remission
296.30 Major depressive disorder recurrent episode unspecified degree
Depression 296.31 Major depressive affective disorder recurrent episode mild degree
296.32 Major depressive affective disorder recurrent episode moderate degree
296.33 Major depressive affective disorder recurrent episode severe degree without
psychotic behavior
296.34 Major depressive affective disorder recurrent episode severe degree specified
as with psychotic behavior
296.35 Major depressive affective disorder recurrent episode in partial or unspecified
remission
296.36 Major depressive affective disorder recurrent episode in full remission
300.4 Dysthymic disorder
309.1 Adjustment reaction with prolonged depressive reaction
311.xx Depressive disorder not elsewhere classified
402.01 Malignant hypertensive heart disease with heart failure
402.11 Benign hypertensive heart disease with heart failure
402.91 Unspecified hypertensive heart disease with heart failure
404.01 Hypertensive heart and chronic kidney disease, malignant, with heart failure
and with chronic kidney stage i-iv, or unspecified
404.03 Hypertensive heart and chronic kidney disease, malignant, with heart failure
and chronic kidney disease state v or end state renal disease
404.13 Hypertensive heart and chronic kidney disease, benign, ...
404.93 Hypertensive heart and chronic kidney disease, unspecified,
404.11 Hypertensive heart and chronic kidney disease, benign, with heart failure and
with chronic kidney stage i-iv, or unspecified
404.91 Hypertensive heart and chronic kidney disease, unspecified, with heart failure
and with chronic kidney stage i-iv, or unspecified
425.1 Hypertrophic obstructive cardiomyopathy
425.4 Other primary cardiomyopathies
Congestive Heart
425.5 Alcoholic cardiomyopathy
Failure
425.7 Nutritional and metabolic cardiomyopathy
425.8 Cardiomyopathy in other diseases classified elsewhere
425.9 Secondary cardiomyopathy unspecified
428.0 Congestive heart failure unspecified
428.1 Left heart failure
428.20 Unspecified systolic heart failure
428.21 Acute systolic heart failure
428.22 Chronic systolic heart failure
428.23 Acute on chronic systolic heart failure
428.30 Unspecified diastolic heart failure
428.31 Acute diastolic heart failure
428.32 Chronic diastolic heart failure
428.33 Acute on chronic diastolic heart failure
428.40 Unspecified combined systolic and diastolic heart failure
428.41 Acute combined systolic and diastolic heart failure
HCPCS codes:
84 Acumen, LLC
Condition Codes
Procedure Codes:
CPT Codes:
66982 Extracapsular cataract removal with insertion of intraocular lens prosthesis, (1-
Cataract stage procedure), manual or mechanical technique (eg, irrigation and aspiration
or phacoemulsification), complex, requiring devices or techniques not generally
used in routine cataract surgery (eg, iris expansion device, suture support for
intraocular lens, or primary posterior capsulorhexis) or performed on patients in
the amblyogenic developmental stage.
66983 Intracapsular cataract extraction with insertion of intraocular lens prosthesis (1
stage procedure),
66984 Extracapsular cataract removal with insertion of intraocular lens prosthesis (1
stage procedure), manual or mechanical technique (eg, irrigation and aspiration
or phacoemulsification)
66840 Removal of lens material; aspiration technique, 1 or more stages
66850 Removal of lens material; phacofragmentation technique (mechanical or
ultrasonic) (eg, phacoemulsification), with aspiration
66852 Removal of lens material; pars plana approach, w/ or wo vitrectomy
66920 Removal of lens material; intracapsular
66930 Removal of lens material; intracapsular, for dislocated lens
66940 Removal of lens material; extracapsular (not 66840, 66850, 66852)
714.0 Rheumatoid Arthritis (RA) - does not include juvenile (714.30)
714.1 Felty's syndrome - RA with splenoadenomegaly and leukopenia
714.2 Other RA with visceral or systemic involvement
Rheumatoid 714.30 Chronic or unspecified polyarticular juvenile RA
Arthritis 714.31 Acute polyarticular juvenile RA
714.32 Pauciarticular juvenile RA
714.33 Monoarticular juvenile RA
714.89 Other specified inflammatory polyarthropathies
353.1, 353.4 Nerve root and plexus disorders
355.0 Lesion of sciatic nerve
720.0 - 720.9 Spondylitis
721.3-721.91 Spondylosis and allied disorders
722.10 Intervertebral disc disorders
722.32
Low Back Pain 722.52
722.73
722.83
722.93
724.02-03 Other and unspecified disorders of back. Excludes: collapsed
vertebra
724.2-.9
51.2 Cholecystectomy
51.21 Other partial cholecystectomy
51.22 Cholecystectomy
51.23 Laparoscopic cholecystectomy
51.24 Laparoscopic partial cholecystectomy
Cholecystectomy
CPT Codes:
86 Acumen, LLC
Condition Codes
47600 Cholecystectomy
47605 Cholecystectomy with cholangiography
47610 Cholecystectomy with exploration of common duct
47612 Cholecystectomy with exploration of common duct; with
choledochoenterostomy
47620 Cholecystectomy with exploration of common duct; with transduodenal
sphincterotomy or sphincteroplasty, with or without cholangiography
DX Codes:
Procedure Codes
85.1-85.19 Biopsy
85.20-85.21 Lumpectomy
Breast Cancer 85.22-85.23 partial mastectomy
40.3 Lymph node dissection
85.33-85.48 Mastectomy
CPT Codes:
Procedure Codes:
CPT Codes:
G0416 Surgical pathology, gross and microscopic examination for prostate needle
Prostate Cancer
saturation biopsy sampling, 1-20 specimens
G0417 Surgical pathology, gross and microscopic examination for prostate needle
saturation biopsy sampling, 21-40 specimens
G0418 Surgical pathology, gross and microscopic examination for prostate needle
saturation biopsy sampling, 41-60 specimens
G0419 Surgical pathology, gross and microscopic examination for prostate needle
saturation biopsy sampling, greater than 60 specimens
55700 Biopsy, prostate; needle or punch, single or multiple, any approach
55705 Biopsy, prostate; incisional, any approach
55706 Biopsies, prostate, needle, transperineal, stereotactic template guided
saturation sampling, including imaging guidance
0137T Biopsy, prostate, needle, saturation sampling for prostate mapping
This methodology links the Surveillance, Epidemiology, and End Results (SEER)
registry data with Medicare claims data to validate the incident cancer cohort definitions used in
the analysis. The SEER program of the National Cancer Institute (NCI) collects data from
hospitals, physicians, laboratory reports and death certificates to determine cancer status in the
United States, including diagnosis date. The SEER registry is recognized as one of most accurate
sources of cancer incidence statistics: The seventeen SEER registries capture approximately 98
percent of breast cancer cases within the SEER domain74 It is not possible to use the SEER-
Medicare linked data for the purposes of this analysis because the registry covers only 28 percent
of the U.S. population and the goal of the analysis is to examine regional variation in costs,
utilization and quality across the U.S.
The general algorithm for determining membership in the cancer cohorts based on the
claims data involves three basic steps: 1) Screen valid claims for the appropriate diagnosis and
74
H. T. Gold and H. T. Do, "Evaluation of three algorithms to identify incident breast cancer in Medicare claims
data," Health Serv Res 42, no. 5 (2007), http://www.ncbi.nlm.nih.gov/pubmed/17850533.
88 Acumen, LLC
procedure codes, and if a specified code combination is present, define the index date as the date
of the claim containing the first condition relevant code, 2) for the Medicaid analysis, require
continuous enrollment in Medicaid without benefits restrictions for the duration of the clean
period, and 3) Remove prevalent cases, or beneficiaries with claims dating from the clean
period that contain procedure and diagnosis codes indicating preexisting cancer. Since the
analysis aims to assess the costs in the first year after a cancer diagnosis, membership in all three
cancer cohorts requires a clean period ranging from 6 months to 2 years. Because Medicaid
beneficiaries tend to enroll and disenroll frequently, a substantial number of beneficiaries are lost
as a result of the continuous enrollment requirement. More specifics about the screening
requirement (step 1) can be found in Appendix A. The cancer cohort algorithms were developed
based on the academic literature and refined under the guidance of the IOM committee member
Peter Bach of the Memorial Sloan-Kettering Cancer Center.
The cancer cohort algorithms are assessed using Medicare data based on four criteria:
positive predictive value, negative predictive value, sensitivity and specificity. Positive
predictive value measures the percentage of beneficiaries assigned by the algorithm to a given
incident cancer cohort that are registered for that cancer incident to the index date defined in step
1 in the SEER registry. Negative predictive value measures the percentage of beneficiaries not
assigned by the algorithm to a given incident cancer cohort that are not registered for that
condition incident to the index date in the SEER registry. Sensitivity measures the probability an
individual is assigned to a given incident cancer cohort given the individual is registered for the
condition incident for the index date in the SEER registry. Specificity measures the probability
an individual is not assigned to a given incident cancer cohort given the individual is not
registered for the condition incident to the index date in the SEER registry. The table below
provides these metrics for the final breast, prostate, and lung cancer algorithms.
Positive Negative
Condition Cohort Sensitivity
Predictive Value Predictive Value
Specificity
Breast Cancer 92.11 99.87 71.22 99.97
Prostate Cancer 83.82 99.84 71.57 99.92
Lung Cancer 72.82 99.86 76.75 99.82
They are not enrolled in Medicare Part A or Part B during the first month of the
observation period (column C)
They have an invalid ZIP code that is missing or does not map to a HRR, HSA, and MSA
(column D)
They are listed in the EDB as the primary payer for their health care costs (column E)
They have claims with third party payer costs in the observation window (column F)
For acute cohorts, they have $0 total cost on all Part A and B claims in the observation
window.
Column A presents the total number of Medicare episodes created in the 2007-2009
period for each condition cohort and for the aggregate sample. The aggregate sample includes
over 114 million episodes, and the condition cohorts range from 151,000 episodes (for breast
cancer) to over 22.8 million episodes (for low back pain). Each of the remaining columns shows
the percent of episodes which satisfy each criterion. Because these exclusions need not be
applied in any order, some episodes will be excluded based on more than one restriction. As a
result, the total percent of episodes excluded is not necessarily equal to the sum of the episodes
lost to each restriction.
90 Acumen, LLC
primary payer must not overlap the observation period. Column F shows the percent of episodes
that do not have any claims with third party payer costs. In general, columns E and F are
comparable, which suggests that about 2 percent of episodes are excluded due to Medicare not
being the primary payer. Column G presents, for acute cohorts, the percent of episodes with $0
total cost on Parts A and B during the observation window. Acute episodes with $0 total cost are
likely miscoded or otherwise not valid episodes of care. Column H presents the final results of
applying these restrictions. The percent of episodes from the total that are ultimately excluded
from the analysis ranges from 4 percent (for prostate cancer) to 25 percent (for CHF).
A B C D E F G H
Not A or B Acute
Enrolled Medicare
Enrolled Claims
Total Not as is Not Total
in First Invalid with $0
Number of Found Primary Primary Episodes
Month of ZIP Code Total
Episodes in EDB Payer Payer for Lost
Observati Cost on A
(EDB) all Claims
on and B
Aggregate 114,778,863 0.0% 0.0% 1.9% 5.8% 2.8% N/A 8.8%
AMI 1,221,791 0.0% 14.1% 0.8% 1.1% 0.8% 0.0% 16.2%
Arthritis 1,911,765 0.0% 8.5% 0.7% 2.6% 2.9% N/A 12.9%
Breast Cancer 151,390 0.0% 6.0% 0.5% 1.9% 2.1% 0.0% 9.2%
Cataract 4,082,817 0.0% 1.9% 0.5% 1.8% 1.9% 0.1% 5.0%
CHD 22,667,833 0.0% 13.7% 0.7% 1.9% 1.9% N/A 16.8%
CHF 13,152,055 0.0% 22.3% 0.6% 1.3% 1.2% N/A 24.5%
Cholecystect-
409,482 0.0% 11.8% 0.9% 1.7% 1.3% 0.0% 14.6%
omy
COPD 13,147,737 0.0% 16.7% 0.5% 1.6% 1.5% N/A 19.2%
Depression 8,685,355 0.0% 9.4% 0.6% 2.2% 2.1% N/A 13.1%
Diabetes 21,478,295 0.0% 10.2% 0.7% 2.2% 2.1% N/A 13.7%
Low Back Pain 22,819,996 0.0% 7.2% 0.5% 2.4% 2.3% N/A 11.0%
Lung Cancer 335,482 0.0% 3.5% 0.3% 2.4% 2.7% 0.0% 7.3%
Pneumonia 2,628,307 0.0% 6.2% 0.6% 1.3% 1.1% 0.0% 8.4%
Prostate Cancer 224,580 0.0% 0.6% 0.6% 2.2% 2.6% 0.0% 4.4%
Stroke 730,743 0.0% 13.4% 1.0% 1.0% 0.8% 0.0% 15.4%
75
"IPPS Annual Proposed and Final Rules," Centers for Medicare & Medicaid Services,
http://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/AcuteInpatientPPS/.
92 Acumen, LLC
(B.3) _ _ _ /
For OT claims without a valid HCPCS, the Medicaid study adjusts costs with the
Geographic Adjustment Factor (GAF). The GAF is a weighted average of the three GPCIs where
the cost share weights are determined by the Medicare Economic Index (MEI) 2006 base year
weights as shown in equation (B.4).76 The methodology uses equation (B.5) to standardize costs
on OT claims with no valid HCPCS code.
(B.5) _ _ _ /
76
"Medicare Economic Index Web Table," Centers for Medicare & Medicaid Services,
http://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-Trends-and-
Reports/MedicareProgramRatesStats/downloads//mktbskt-economic-index.pdf.
77
"Berenson-Eggers Type of Service (BETOS)," Centers for Medicare & Medicaid Services,
https://www.cms.gov/Medicare/Coding/HCPCSReleaseCodeSets/BETOS.html.
95
Medicare /
Acumen, LLC
Service
Medicaid Medicare Specification Medicaid Specification
Category
Claim Types
(nursing facility), 34 (hospice), 51 (inpatient psychiatric), 52 Type of Service in (7 (nursing facility), 13 (home
(psychiatric facility), 53 (community mental health center), health), 33 (rehab), 35 (hospice))
56 (psychiatric residential treatment center), 61 (inpatient
rehab), 62 (outpatient rehab))
OP, PB /
Procedures OP and PB claims: first digit of BETOS="P" (procedures) OT claims: first digit of BETOS="P" (procedures)
OT
OP claims: BETOS = "O1A" (ambulance) OR Revenue OT claims: BETOS = O1A (ambulance) OR
Emergency Center code 0450-0459 or 0981 (emergency room) Revenue Center code 0450-0459, or 0981 (emergency
OP, PB /
Room / PB claims: BETOS = "O1A" (ambulance) OR Place of room) OR Place-of-Service = 23 (ambulance)
OT, IP
Ambulance Service IN (23 (emergency room), 41 (land ambulance), 42 IP claims: Revenue Center code 0450-0459 or 0981
(air or water ambulance)) (emergency room)
All claim
Other Any claim not previously categorized Any claim not previously categorized
types
The Medicare analysis further divides post-acute care into categories to determine the main cause of variation. The table below
presents the substratifications of the post-acute care categories.
IOM Study of Geographic Variation | May 2013
Post-Acute Care
Medicare Specification
Category
IP claims with last four digits of Provider Number: 2000-2299 (long term
care)
PB claims: Place of Service in: 51 (inpatient psychiatric), 52 (psychiatric
facility), 53 (community mental health center), 56 (psychiatric residential
treatment center), 61 (inpatient rehab), 62 (outpatient rehab))
Medicare Method
Acumen, LLC
RX claims. (Under-
For each person/NDC/Day: =1 if days claims with HCPCs
All valid NDC reported in states with
Number of RX drug fills NDC supply <=30; =days supply/30 if days with BETOS in (D1G,
claims prevalent local code
supply >30 O1E, or O1D), and all
system usage)
Part D claims
99281 CPT Emergency dept visit, minimal OP claims with
99282 CPT Emergency dept visit, minor Revenue Center code
99283 CPT Emergency dept visit, low 0450-0459 or 0981; OT claims with Revenue
99284 CPT Emergency dept visit, moderate and IP claims with Center code 0450-0459 or
Revenue Center code 0981, or Place-of-Service
Number of Emergency
0450-0459 or 0981 = 23; and IP claims with
Department Visit Days
and restricting to Revenue Center code
99285 CPT Emergency dept visit, high Source of Admission 0450-0459 or 0981. Max 1
=7. per person per day.
Max 1 per person per
day.
98
Medicare Method
Utilization Count Measure Code Medicaid Method
Code Description Notes
type
Maximum of 1 OT claims (Under-
Number of Imaging Diagnostic procedure of each type reported in states with
CPT
Encounters Imaging Codes on any given day. OP prevalent local code
and PB claims. system usage.)
OT and IP claims (Under-
No more than 1 test
Cardiac Stress ICD-9 reported in states with
Sentinel Services Nuclear stress tests per enrollee day. IP,
Test and CPT prevalent local code
OP, and PB claims.
system usage.)
No more than 1 OT and IP claims (Under-
Bilateral Cardiac ICD-9 procedure per enrollee reported in states with
Sentinel Services Bilateral cardiac catheterization
Catheterization and CPT day. IP, OP, and PB prevalent local code
claims. system usage.)
Medicare Method
Acumen, LLC
78
The analysis uses CMS definition of physician specialty code found in the Medicare claims processing manual: "Medicare Claims Processing Manual:
Chapter 26 - Completing and Processing Form CMS-1500 Data Set. Section 10.8.2: Physician Specialty Codes," The Centers for Medicare & Medicaid Services,
http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/clm104c26.pdf.
100
79
At the recommendation of the team clinician, the chosen approach enforces all optional exclusions in the quality measures in order to give providers the benefit
of the doubt in cases where certain services may not be appropriate for beneficiaries with certain characteristics.
101
79
Acumen, LLC
Pneumonia Bacterial Pneumonia Outcome Programmed to AHRQ quality indicators. Guide to prevention quality http://www.qualityindica
Admission Rate: All align with AHRQ indicators: hospital admission for ambulatory care sensitive tors.ahrq.gov/Downloads
discharges of age 18 PQI #11 conditions [version 3.1]. Rockville (MD): Agency for /Software/SAS/V41A/Te
years and older with specifications Healthcare Research and Quality (AHRQ); 2007 Mar 12. chSpecs/PQI%2011%20
ICD-9-CM principal 59 p. (AHRQ Pub; no. 02-R0203). Bacterial%20Pneumonia
diagnosis code for %20Admission%20Rate.
bacterial AHRQ quality indicators. Prevention quality indicators pdf
pneumonia. appendices [version 4.2]. Rockville (MD): Agency for
Healthcare Research and Quality (AHRQ); 2010 Sep. 1 p.
79
Condition Quality Measure Type Specification External Source Link to measure
Rheumatoi Rheumatoid Process Programmed to National Committee for Quality Assurance (NCQA). See HEDIS measure:
d Arthritis arthritis: Percentage align with HEDIS 2011: Healthcare Effectiveness Data and Rheumatoid arthritis:
of members who HEDIS Information Set. Vol. 1, narrative. Washington (DC): percentage of members
were diagnosed with specifications National Committee for Quality Assurance (NCQA); 2010. who were diagnosed with
rheumatoid arthritis various pages. rheumatoid arthritis and
and who were who were dispensed at
dispensed at least National Committee for Quality Assurance (NCQA). least one ambulatory
one ambulatory HEDIS 2011: Healthcare Effectiveness Data and prescription for a disease
prescription for a Information Set. Vol. 2, technical specifications. modifying anti-rheumatic
disease modifying Washington (DC): National Committee for Quality drug (DMARD).
anti-rheumatic drug Assurance (NCQA); 2010. various pages.
(DMARD).
Depression Antidepressant Process Programmed to National Committee for Quality Assurance (NCQA). See HEDIS measure:
medication align with HEDIS 2011: Healthcare Effectiveness Data and Antidepressant
management HEDIS Information Set. Vol. 1, narrative. Washington (DC): medication management
(effective acute specifications National Committee for Quality Assurance (NCQA); 2010. (effective acute phase
phase treatment): various pages. treatment): percentage of
Percentage of members 18 years of age
members 18 years of National Committee for Quality Assurance (NCQA). and older who were
age and older who HEDIS 2011: Healthcare Effectiveness Data and diagnosed with a new
were diagnosed with Information Set. Vol. 2, technical specifications. episode of major
a new episode of Washington (DC): National Committee for Quality depression, and treated
major depression, Assurance (NCQA); 2010. various pages. with antidepressant
and treated with medication, and who
antidepressant remained on an
medication, and who antidepressant
remained on an medication for at least 84
antidepressant days (12 weeks)
medication for at
least 84 days (12
weeks)
Antidepressant Process Programmed to National Committee for Quality Assurance (NCQA). See HEDIS measure:
medication align with HEDIS HEDIS 2011: Healthcare Effectiveness Data and Antidepressant
management specifications Information Set. Vol. 1, narrative. Washington (DC): medication management
Acumen, LLC
(effective National Committee for Quality Assurance (NCQA); 2010. (effective continuation
continuation phase various pages. phase treatment):
treatment): percentage of members
Percentage of National Committee for Quality Assurance (NCQA). 18 years of age and older
103
79
Acumen, LLC
age and older during various pages. 18 years of age and older
the measurement during the measurement
year who were National Committee for Quality Assurance (NCQA). year who were
hospitalized and HEDIS 2011: Healthcare Effectiveness Data and hospitalized and
discharged alive Information Set. Vol. 2, technical specifications. discharged alive from
from July 1 of the Washington (DC): National Committee for Quality July 1 of the year prior to
year prior to the Assurance (NCQA); 2010. various pages. the measurement year to
measurement year to June 30 of the
June 30 of the measurement year with a
measurement year diagnosis of AMI and
with a diagnosis of who received persistent
AMI and who beta-blocker treatment
received persistent for six months after
beta-blocker discharge
treatment for six
104
79
Condition Quality Measure Type Specification External Source Link to measure
months after
discharge
Coronary Antiplatelet Process Programmed to American College of Cardiology, American Heart Measure 6 (Page 55):
Heart Therapy: Percentage align with Association, Physician Consortium for Performance http://www.ama-
Disease of patients aged 18 ACC/AHA/ PCPI Improvement. Clinical performance measures: chronic assn.org/ama1/pub/uploa
and older with a Chronic Stable stable coronary artery disease. Tools developed by d/mm/pcpi/cadminisetjun
diagnosis of Coronary Artery physicians for physicians. Chicago (IL): American Medical e06.pdf
coronary artery Disease Association (AMA); 2005. 8 p.
disease seen within Performance
a 12 month period Measure #6
who were prescribed specifications
aspirin or
clopidogrel
COPD Chronic Obstructive Outcome Programmed in AHRQ quality indicators. Guide to prevention quality http://www.qualityindica
Pulmonary Disease accordance with indicators: hospital admission for ambulatory care sensitive tors.ahrq.gov/Downloads
(COPD) Admission AHRQ's PQI #5 conditions [version 3.1]. Rockville (MD): Agency for /Software/SAS/V41A/Te
Rate specifications Healthcare Research and Quality (AHRQ); 2007 Mar 12. chSpecs/PQI%2005%20
59 p. (AHRQ Pub; no. 02-R0203). Chronic%20Obstructive
%20Pulmonary%20Dise
AHRQ quality indicators. Prevention quality indicators ase%20%28COPD%29
appendices [version 4.2]. Rockville (MD): Agency for %20Admission%20Rate.
Healthcare Research and Quality (AHRQ); 2010 Sep. 1 p. pdf
Acumen, LLC
79
Acumen, LLC
79
Condition Quality Measure Type Specification External Source Link to measure
Low Back Use of imaging Process Programmed to National Committee for Quality Assurance (NCQA). See HEDIS measure:
Pain studies for low back align with HEDIS 2011: Healthcare Effectiveness Data and Use of imaging studies
pain: Percentage of HEDIS Information Set. Vol. 1, narrative. Washington (DC): for low back pain:
members with a specifications National Committee for Quality Assurance (NCQA); 2010. percentage of members
primary diagnosis of various pages. with a primary diagnosis
low back pain who of low back pain who did
did not have an National Committee for Quality Assurance (NCQA). not have an imaging
imaging study (plain HEDIS 2011: Healthcare Effectiveness Data and study (plain x-ray, MRI,
x-ray, MRI, CT Information Set. Vol. 2, technical specifications. CT scan) within 28 days
scan) within 28 days Washington (DC): National Committee for Quality of the diagnosis
of the diagnosis Assurance (NCQA); 2010. various pages.
Cholecyste Laparoscopic Process Programmed to AHRQ quality indicators. Guide to inpatient quality http://www.qualityindica
ctomy Cholecystectomy align with AHRQ indicators: quality of care in hospitals - volume, mortality, tors.ahrq.gov/Downloads
Rate Inpatient Quality and utilization [version 3.1]. Rockville (MD): Agency for /Software/SAS/v41A/Te
Indicator (IQI) Healthcare Research and Quality (AHRQ); 2007 Mar 12. chSpecs/IQI%2023%20L
Measure #23 91 p. aparoscopic%20Cholecy
specifications stectomy%20Rate.pdf
AHRQ quality indicators. Inpatient quality indicators:
technical specifications [version 4.2]. IQI #23 laparoscopic
cholecystectomy rate. Rockville (MD): Agency for
Healthcare Research and Quality (AHRQ); 2010 Sep. 1 p.
Breast Percentage of Process Programmed to National Committee for Quality Assurance (NCQA). See HEDIS measure:
Cancer women 42 to 69 align with HEDIS 2011: Healthcare Effectiveness Data and Breast cancer screening:
years of age who HEDIS Information Set. Vol. 1, narrative. Washington (DC): percentage of women 40
had one or more specifications, National Committee for Quality Assurance (NCQA); 2010. to 69 years of age who
mammograms modifying age various pages. had one or more
during the range as defined mammograms during the
measurement year for measure. measurement year or the
or the year prior to year prior to the
the measurement measurement year.
year.
Acumen, LLC
107
79
Acumen, LLC
80
"Quality Indicator User Guide: Patient Safety Indicators (PSI) Composite Measures, Version 4.4," The Agency for Healthcare Research and Quality,
Acumen, LLC
http://qualityindicators.ahrq.gov/Downloads/Modules/PSI/V44/Composite_User_Technical_Specification_PSI%20V4.4.pdf.
81
"Quality Indicator User Guide: Inpatient Quality Indicators (IQI) Composite Measures, Version 4.4," The Agency for Healthcare Research and Quality,
http://qualityindicators.ahrq.gov/Downloads/Modules/IQI/V44/Composite_User_Technical_Specification_IQI%20V4.4.pdf.
82
"Quality Indicator User Guide: Prevention Quality Indicators (PQI) Composite Measures, Version 4.4," The Agency for Healthcare Research and Quality,
http://qualityindicators.ahrq.gov/Downloads/Modules/PQI/V44/Composite_User_Technical_Specification_PQI%20V4.4.pdf.
109
Age X X X X X X X X X
Sex X X X X X X X X X
Age-Sex Interaction X X X X X X X X X
Health Status X X X X X
Race X X X X X
85
Income X X X
Institutionalization Status X X
New Enrollee Indicator X X X X X
Dual Enrollment Status X X
Supplemental Medicare Insurance X X X
Hospital Competition X X X
Percent of Population Uninsured X X X
Market-Level Variables
83
Cluster 6 is not applicable to the Medicare or Medicaid analysis.
84
The indicator variable for year of analysis is only used in the analysis of the full 2007 through 2009 data.
85
Cluster 8 does not include the income indicator because the income indicator is highly collinear with the dual enrollment status indicator.
110
86
Cluster 4 is the same as cluster 1 for the Medicaid analysis.
87
Cluster 6 is not applicable to the Medicare or Medicaid analysis.
88
The indicator variable for year of analysis is only used in the analysis on the full 2007 through 2009 data.
111
89
When risk-adjusting the composite quality measures for the aggregate analysis, the regression uses the HCCs from the prior calendar year instead of the HCCs
from the year prior to the inpatient event to assure consistency across the components of the quality measures.
112
90
Acumen, LLC
For beneficiaries in the cholecystectomy cohort whose observation period overlaps 2006, because 2006 is the earliest year of data available for this analysis,
the institutionalization status indicator for both the Medicare and the Medicaid analysis examines the entire 2006 time period.
91
Because the observation window for the cataract cohort can vary between three and six months, the fourth through sixth indicators of the partial year
enrollment variable for cataracts are zero if either the beneficiary is not enrolled in the relevant part of Medicare for that month or the observation window has
ended.
C.4 Market-Level Characteristics
Market-Level Data
Variables
Variable Source
Herfindahl index (HHI) of competition based on the distribution of beds
in each market
Hospital Teaching hospital=1 if there is at least 1 teaching hospital in the HRR
AHA
Competition Specialty hospital=1 if there is at least 1 specialty hospital in the HRR
Government-owned hospital=1 if there is at least 1 government owned
hospital in the HRR
Percent of
Population InterStudy Population uninsured
Uninsured
Supply of Medical Number of hospitals per 1,000
ARF
Services Number of hospital beds per 1,000
Malpractice
MPFS Medicare malpractice GPCI
Environment Risk
Physicians per 1,000
Physician
ARF Active primary care physicians per 1,000
Composition
Active specialists per 1,000
Indicator for Health Professional Shortage Areas (HPSAs), weighted by
Access to Care ARF
county population.
Medicare analysis only: percent of Medicare population covered by
managed care plans
Payer Mix InterStudy
Medicaid analysis only: percent of Medicaid population covered by
managed care plans
Medicaid
InterStudy = (# Medicaid beneficiaries / total population in HRR)
Penetration
Six variables for non-physicians per capita. These variables are not
included in Harvards market-level file. This analysis has created these
variables from ARF data.
Physicians Assistants
Health Professional
ARF Active Nurse Practitioners
Mix
Nurse Anesthetists
Active Certified Nurse Midwives
Registered Nurses
Licensed Practical Nurses and Licensed Vocational Nurses
Medicare analysis only:
Percent of beneficiaries with supplemental Medicare insurance:
(Number of beneficiary-months enrolled in supplementary Medicare
insurance) / (Number of beneficiary-months alive and enrolled in
Percent of Medicare
Medicare Part A or B but not C). Beneficiaries are counted for each
Beneficiaries with
EDB month.
Supplemental
Insurance
This indicator as defined on claims includes Medicaid enrollment.
However, because this analysis includes a dual-enrollment indicator,
beneficiaries enrolled in Medicaid are not counted as having
supplemental insurance.
Table D.14: Medicare Service Utilization Within and Across Regions, Stroke
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $4,513 $4,498 $3,718
Weighted Average of HRR Standard Deviations $4,760 $4,692 $3,869
Standard Deviation of HRR Means $681 $595 $435
Average of HRR 90-10 Differences $9,166 $9,190 $7,486
Weighted Average of HRR 90-10 Differences $9,620 $9,542 $7,742
90-10 Difference of HRR Means $1,593 $1,498 $1,109
Table D.15: Medicare Service Utilization Within and Across Regions, CHD
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $2,862 $2,847 $2,441
Weighted Average of HRR Standard Deviations $3,008 $2,953 $2,512
Standard Deviation of HRR Means $232 $213 $137
Average of HRR 90-10 Differences $4,597 $4,609 $4,018
Weighted Average of HRR 90-10 Differences $4,825 $4,775 $4,127
90-10 Difference of HRR Means $552 $485 $352
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs . 0.15 0.02 0.05 0.13 0.04 0.16 0.05
Acute Care 0.15 1.00 0.00 0.02 0.13 0.00 0.13 0.03
Prescription Drugs 0.02 0.00 1.00 0.24 -0.04 0.29 0.08 0.08
Diagnostic 0.05 0.02 0.24 1.00 -0.05 0.30 0.20 0.13
Post-Acute Care 0.13 0.13 -0.04 -0.05 1.00 0.00 0.07 0.03
Procedures 0.04 0.00 0.29 0.30 0.00 1.00 0.07 0.09
ER/Ambulance 0.16 0.13 0.08 0.20 0.07 0.07 1.00 0.07
Other 0.05 0.03 0.08 0.13 0.03 0.09 0.07 1.00
Table D.18: Stroke Medicare Service Category Utilization, Pearson Correlation (2007)
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs . 0.24 0.05 0.12 0.22 0.06 0.26 0.08
Acute Care 0.24 1.00 0.02 0.07 0.22 0.01 0.23 0.04
Prescription Drugs 0.05 0.02 1.00 0.20 -0.02 0.17 0.03 0.06
Diagnostic 0.12 0.07 0.20 1.00 -0.04 0.36 0.10 0.10
Post-Acute Care 0.22 0.22 -0.02 -0.04 1.00 0.00 0.15 0.04
Procedures 0.06 0.01 0.17 0.36 0.00 1.00 0.04 0.10
ER/Ambulance 0.26 0.23 0.03 0.10 0.15 0.04 1.00 0.04
Other 0.08 0.04 0.06 0.10 0.04 0.10 0.04 1.00
Table D.20: Diabetes Medicare Service Category Utilization, Pearson Correlation (2007)
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Breast 0.587 0.551 0.311 0.406 0.530 0.501 0.475 0.516 0.583 0.560 0.527 0.537 0.574 1.000 0.608 0.524
Cancer Lung 0.551 0.679 0.339 0.488 0.674 0.588 0.433 0.537 0.622 0.595 0.496 0.535 0.556 0.608 1.000 0.519
Prostate 0.501 0.545 0.230 0.409 0.548 0.524 0.373 0.427 0.502 0.467 0.454 0.478 0.485 0.524 0.519 1.000
D.6 Variation in Medicare Quality of Care across Cohorts
122
Regions that perform well on one quality metric for Medicare beneficiaries do not necessarily perform well on another quality
metric. Table D.22 displays the correlation between the composite quality measures, and Table D.23 displays the correlation between
quality of care as measured by each condition-specific quality metric.92 The correlations between the composite quality measures
ranges from -0.02 (for the PQI and PSI measures) to 0.24 (for the PSI and IQI measures). Mechanically, the potentially avoidable
complications and iatrogenic events measured by the composite PSI may result in an inpatient mortality that is captured by the IQI,
which may cause the positive correlation between quality of care for the PSI and IQI measures. The correlations between the
condition-specific measures, shown in Table D.23, range from -0.38 (for the diabetes retinal screening measure and the
cholecystectomy measure) to 0.90 (for the depression acute phase treatment measure and the depression chronic phase treatment
measure). If one treats each condition-specific measure-to-measure correlation as a single observation, the average correlation
between an HRRs quality score on one measure and its quality score on any other measure is 0.07. Intuitively, having high-quality
providers for one condition in a region does not necessarily suggest that the quality of care for another condition is also above average.
For example, while the Lake Charles, LA, HRR is the highest-performing HRR for the cataract measure, it is in the worst-performing
10th percentile for the cholecystectomy measure; cataract surgery and cholecystectomies are performed by very different types of
physicians.
Table D.22: Pearson Correlation between Composite Quality Measures for Medicare Beneficiaries
PSI IQI PQI
PSI 1.00
IQI 0.24 1.00
PQI -0.02 0.18 1.00
Acumen, LLC
92
For some quality measures, a lower score indicates better quality. The correlations in both tables have been renormalized for interpretability so that a positive
correlation always means that higher costs are associated with a higher quality of care. In addition, both tables shows Pearson correlations, and the Spearman
rank correlations show similar results.
123
Table D.23: Pearson Correlation between Condition-Specific Quality Measures for Medicare Beneficiaries
Acumen, LLC
These results must be interpreted with caution and not as evidence that increased
spending causes higher or lower quality outcomes for two reasons. First, beneficiaries who are
included in a disease cohort are sicker than the general population by default, and the risk
adjustment methodology for the quality measures that are risk adjusted may not adequately
capture these differences in health status. Second, mechanical relationships between the quality
measures and utilization may cause the correlations to be artificially strong. For example, the
outcome in the COPD admissions measure is an inpatient admission. As the rate of admissions
increases in a region, indicating a lower quality of COPD care, spending in the region will also
increase because admissions are high-cost. Thus, this relationship may induce a negative
correlation between spending and the quality of care provided.
93
For some quality measures, a lower score indicates better quality. The correlations in Table D.24 have been
renormalized for interpretability so that a positive correlation always means that higher costs are associated with a
higher quality of care. In addition, while Table D.24 presents Pearson correlations, the Spearman rank correlations
show similar results.
Table E.14: Stroke Dispersion of Medicaid Service Utilization Within and Across Regions
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $14,194 $14,211 $13,097
Weighted Average of HRR Standard Deviations $15,573 $15,418 $14,228
Standard Deviation of HRR Means $1,804 $1,780 $1,377
Average of HRR 90-10 Differences $8,771 $8,806 $8,263
Weighted Average of HRR 90-10 Differences $9,030 $8,933 $8,277
90-10 Difference of HRR Means $4,359 $4,274 $3,204
Table E.15: CHD Dispersion of Medicaid Service Utilization Within and Across Regions
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $12,078 $12,120 $11,215
Weighted Average of HRR Standard Deviations $12,284 $12,198 $11,032
Standard Deviation of HRR Means $1,211 $1,205 $985
Average of HRR 90-10 Differences $6,623 $6,680 $6,045
Weighted Average of HRR 90-10 Differences $6,270 $6,221 $5,361
90-10 Difference of HRR Means $3,003 $3,009 $1,943
Table E.16: Diabetes Dispersion of Medicaid Service Utilization Within and Across Regions
Price-
Price-
Unadjusted Standardized
Standardized
Risk-Adjusted
Average of HRR Standard Deviations $8,763 $8,832 $7,783
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs .
Acute Care -0.08 1.00
Prescription Drugs 0.04 -0.01 1.00
Diagnostic 0.40 0.27 0.08 1.00
Post-Acute Care 0.05 0.03 0.08 0.03 1.00
Procedures 0.31 0.25 0.12 0.38 0.03 1.00
ER/Ambulance -0.12 -0.15 0.04 0.27 0.01 0.11 1.00
Other 0.01 -0.01 0.12 0.03 -0.01 0.07 0.01 1.00
Table E.18: Stroke Medicaid Service Category Utilization, Pearson Correlation (2007)
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs .
Acute Care -0.04 1.00
Prescription Drugs 0.06 0.00 1.00
Diagnostic 0.42 0.30 0.09 1.00
Post-Acute Care 0.08 0.05 0.11 0.06 1.00
Procedures 0.20 0.18 0.10 0.32 0.00 1.00
ER/Ambulance -0.08 -0.13 0.03 0.25 0.06 0.04 1.00
Other 0.01 0.00 0.14 0.03 -0.03 0.06 0.01 1.00
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs .
Acute Care 0.02 1.00
Prescription Drugs 0.05 0.00 1.00
Diagnostic 0.36 0.23 0.11 1.00
Post-Acute Care 0.04 0.03 0.05 0.01 1.00
Procedures 0.28 0.22 0.12 0.36 0.01 1.00
ER/Ambulance -0.02 -0.06 0.02 0.28 0.03 0.12 1.00
Other 0.03 0.01 0.13 0.06 0.00 0.07 0.01 1.00
Table E.20: Diabetes Medicaid Service Category Utilization, Pearson Correlation (2007)
Prescription Drugs
Post-Acute Care
ER/Ambulance
Remaining Costs
Acute Care
Procedures
Diagnostic
Other
Remaining Costs .
Acute Care 0.11 1.00
Prescription Drugs 0.09 0.02 1.00
Diagnostic 0.39 0.27 0.12 1.00
Post-Acute Care 0.06 0.04 0.05 0.02 1.00
Procedures 0.21 0.13 0.12 0.30 0.02 1.00
ER/Ambulance 0.12 0.06 0.04 0.35 0.04 0.12 1.00
Other 0.06 0.01 0.12 0.06 0.00 0.06 0.03 1.00
94
For some quality measures, a lower score indicates better quality. The correlations in both tables have been renormalized for interpretability so that a positive
correlation always means that higher costs are associated with a higher quality of care. In addition, both tables show Pearson correlations, and the Spearman rank
correlations show similar results.
133
suggesting there is little relationship between a regions overall quality of care for one metric and the regions quality of care for
Acumen, LLC
another metric for Medicaid beneficiaries. Table E.23 displays the correlation between each condition-specific quality metric. In Table
E.23, results are not shown for the breast cancer radiation, breast cancer screening, or stroke quality measures because too few HRRs
had a sufficient number of events in the denominator of the measure to be reported. The correlations between the condition-specific
measures range from -0.334 (for the COPD admissions and diabetes retinal screening measures) to 0.744 (for the two depression
measures). If each measure-to-measure correlation is treated as a single observation, the average correlation between an HRRs quality
score on one measure and its quality score on any other measure for Medicaid beneficiaries is 0.06. Intuitively, because different types
of providers treat different conditions, having providers that perform high-quality care for one condition in a region does not
necessarily imply that the region also has providers that perform above-average for another condition.
Table E.22: Pearson Correlation between Composite Quality Measures for Medicaid Beneficiaries
PSI IQI PQI
PSI 1.00
IQI 0.041 1.00
PQI -0.198 -0.131 1.00
IOM Study of Geographic Variation | May 2013
Table E.23: Pearson Correlation between Condition-Specific Quality Measures for Medicaid Beneficiaries
134
AMI
Arth. 0.413
BC Rad
BC Screen
Cat. 0.077 0.154
CHD -0.072 0.109 0.040
CHF 0.359 -0.095 0.015 -0.018
Chol. -0.253 0.001 -0.106 -0.109 0.108
COPD
0.319 0.039 0.024 -0.052 0.072 0.168
Adm.
COPD
0.706 0.208 -0.011 0.099 0.230 0.042 0.264
Bronch.
Depr.
0.277 0.293 0.188 0.065 -0.050 0.027 0.069 0.128
Acute
Depr.
0.361 0.238 0.196 0.072 -0.067 -0.008 -0.070 0.076 0.744
Chron.
Diab. Amp. 0.495 0.151 -0.111 0.000 0.252 0.117 0.012 0.348 -0.171 -0.090
Diab.
-0.103 -0.016 0.187 -0.102 0.108 -0.160 -0.170 -0.182 -0.032 0.049 -0.093
Hemo.
Diab. Ret. -0.268 -0.053 0.236 0.015 0.182 -0.114 -0.334 -0.145 -0.057 -0.013 -0.115 0.727
LBP 0.120 -0.020 -0.108 0.045 -0.032 0.048 -0.029 -0.025 -0.016 0.076 0.011 -0.156 -0.238
Pneu. -0.047 -0.146 -0.095 -0.015 0.556 0.163 -0.016 0.064 -0.150 -0.133 0.240 0.199 0.228 0.116
Str.
Acumen, LLC
E.7 Relationship between Medicaid Utilization and Quality of Care
Regions with high utilization do not necessarily provide a higher quality of care for
Medicaid beneficiaries. Table E. 24 shows the correlation between HRR-level quality indices
and price-standardized, risk adjusted costs for each quality measure. No results are shown for the
breast cancer radiation, breast cancer screening, or stroke quality measures because too few
HRRs had a sufficient number of events in the denominator of the measure to be reported. In
Table E.24, positive correlations, which are highlighted in red, indicate that higher spending on
patients in that cohort is associated with a higher quality of care for that measure.95 The average
of the correlations is weakly negative, at -0.031. The strongest positive correlation with
utilization exists for the composite PSI measure (0.24), and the strongest negative correlation
with utilization exists for the diabetes retinal screening measure (-0.267)
These results must be interpreted with caution and not as evidence that increased
spending causes higher or lower quality outcomes for two reasons. First, beneficiaries who are
included in a disease cohort are sicker than the general population by default, and the risk
adjustment methodology for the quality measures that are risk adjusted may not adequately
capture these differences in health status. Second, some mechanical relationships between the
quality measures and utilization may cause the correlations to be artificially strong. For example,
the outcome in the CHF admissions measure is an inpatient admission. As the rate of admissions
increases in a region, indicating a lower quality of CHF care, spending in the region will also
increase because admissions are high-cost. Thus, this relationship may induce a negative
correlation between spending and the quality of care provided.
95
For some quality measures, a lower score indicates better quality. The correlations in Table E.24 have been
renormalized for interpretability so that a positive correlation always means that higher costs are associated with a
higher quality of care. In addition, while Table E.24 presents Pearson correlations, the Spearman rank correlations
show similar results.
1. How much variation is there in per capita spending in Medicare Advantage across the
nation?
2. Are regions with high Medicare Advantage premiums likely to have high Medicare
Advantage premiums in subsequent years?
4. Are regions with high Medicare Advantage spending levels also likely to have high
spending levels in Medicare fee-for-service?
The remainder of this appendix proceeds as follows. First, Section F.1 briefly describes CMS
payment policy for Medicare Advantage beneficiaries. Section F.2 describes the methodology
used to answer the five research questions posed above, including the data sources, cohort
definitions, outcome variables, and risk adjustment specifications. Finally, Section F.3 directly
evaluates each research question.
96
This analysis examines all Part C beneficiaries, which include those enrolled in certain private health plans,
known as cost plans, which are not technically MA plans. For simplicity, from this point forward, this report
includes cost plans when referring to Medicare Advantage unless otherwise noted.
97
Centers for Medicare & Medicaid Services, "Medicare Managed Care Manual, Chapter 1: Introduction,"
http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/mc86c01.pdf.
There are five main types of health plans available to Medicare beneficiaries, including
local coordinated care, private fee-for-service, regional preferred provider organization, medical
savings account, and cost plans. First, local coordinated care plans (CCPs) include local health
maintenance organizations (HMOs) and preferred provider organizations (PPOs). Local CCPs
contract with a network of providers and are required to offer at least one plan that includes Part
D drug coverage in their service area. CCPs also generally require a primary care gatekeeper to
refer beneficiaries for certain services. Second, private fee-for-service plans (PFFS) are not
required to establish networks of providers or offer prescription drug coverage. Third, regional
PPOs are similar to local PPOs but provide coverage for an entire geographic region that is
defined by CMS.100 Fourth, medical savings account plans (MSAs) offer a high-deductible plan
with an MSA that beneficiaries can use to cover cost-sharing or non-covered services. Fifth, cost
plans are compensated based on calculated reasonable costs for the Medicare-covered services
they provide for each beneficiary instead of a fixed rate per beneficiary. Cost plans are not
technically MA plans as Congress authorized their creation under a different section of the Social
Security Act. 101Unlike traditional FFS, beneficiaries in cost plans are overseen by a network of
providers, though CMS does pay FFS rates when beneficiaries in cost plans receive services
outside of their network. In 2007, at the start of this studys analysis period, 70 percent of MA
98
"Medicare: Medicare Advantage Fact Sheet," The Henry J. Kaiser Family Foundation,
http://www.kff.org/medicare/upload/2052-15.pdf.
99
"The Medicare Advantage Program: Status Report," MedPAC Report to the Congress: Medicare Payment Policy,
http://www.medpac.gov/chapters/Mar12_Ch12.pdf.
100
CMS sets benchmarks for regional PPOs at the regional level differently than for county-based plans. Regional
benchmarks are a blend of a plan-bid component, based on the weighted average of the bids for regional plans in
that region, and a statutory component, consisting of the weighted average of the county capitation rates in that
region.
101
"Medicare Managed Care Manual, Chapter 1: Section 1876 Cost Plans," Centers for Medicare & Medicaid
Services, http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/mc86c01.pdf.
2. The MAO places a bid for the plans expected monthly costs per beneficiary;
3. The plans bid is compared to the local county benchmark, and the lower amount is
the set as the plans base rate;
4. CMS pays the plan the base rate for each enrollee, adjusted for the plans beneficiary
case mix and including a rebate if the bid is below the benchmark;
MA payments are based first on the capitated rate, or benchmark, CMS sets each year for
MA plans in a given county to cover a minimum set of services per beneficiary. The benchmark
represents the maximum amount that CMS will pay any MA plan to provide this core set of
services. Benchmarks are historically based on the estimated average Medicare FFS costs in that
102
Mark Merlis, "Medicare Advantage Payment Policy," National Health Policy Forum (2007).
Second, for each plan it offers, the MAO places a bid that represents the expected costs
for the plan to cover the core set of services for an average beneficiary, including administrative
fees. The plans bid does not take into account the expected enrollee case mix. CMS may
negotiate bid amounts with most plans.
Third, the plans bid is compared to the local benchmark. The benchmark represents
CMS maximum monthly payment per beneficiary. Thus, if the bid is less than the benchmark,
the base rate is set at the bid; if the bid is greater than or equal to the benchmark, the base rate is
set at the benchmark.
Fourth, to calculate the payment to the plan, the base rate is adjusted for the plans
expected beneficiary case mix using the CMS-HCC model.105 The CMS-HCC model is
estimated using FFS claims for beneficiaries in Original Medicare to predict costs based on
demographic and health factors. Because MA beneficiaries tend to be healthier than beneficiaries
in Original Medicare, the aggregate payments to MA plans after accounting for beneficiary
health status would be lower than if MA beneficiaries represented the average set of beneficiaries
on which the CMS-HCC model was developed. Thus, CMS also multiplies all base rates by a
single budget neutrality adjustment factor to prevent a decrease in payments to MA plans due to
lower risk scores.106 If the plans bid is less than the benchmark, CMS pays the plans the bid and
a rebate for part of the difference between the benchmark and bid after adjusting for health
status. CMS retains 25 percent of the difference, and the plan uses the remaining 75 percent to
cover additional benefits or reduce premiums or cost sharing paid by beneficiaries.
103
Some benchmarks in rural areas were initially set higher than FFS costs to increase plan availability.
"The Medicare Advantage Program," MedPAC Report to the Congress: Medicare Payment Policy,
http://www.medpac.gov/chapters/Mar09_Ch03.pdf.
104
The rebased rates are calculated using a rolling five-year average of FFS costs. "Medicare Managed Care
Manual, Chapter 8: Determination of Annual Capitation Rates," Centers for Medicare & Medicaid Services,
http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/mc86c08.pdf.
105
"Medicare Advantage Rates & Statistics: Risk Adjustment," Centers for Medicare & Medicaid Services,
https://www.cms.gov/Medicare/Health-Plans/MedicareAdvtgSpecRateStats/Risk_adjustment.html.
106
CMS began to phase out the budget neutrality adjustment in 2007, and it was completely phased out by 2011.
"Medicare Managed Care Manual, Chapter 8: Budget Neutral (BN) Risk Adjustment," Centers for Medicare &
Medicaid Services, http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/mc86c08.pdf.
The remainder of this section details the methodology used by this analysis to measure
regional variation in spending. Section F.2.1 describes the data sources used in this study, and
Section F.2.2 defines which beneficiaries are included in the analysis and explains how this study
creates condition-specific beneficiary cohorts. Section F.2.3 specifies how this study calculates
health care expenditures for MA beneficiaries. Finally, Section F.2.4 defines the risk adjustment
econometric specifications. Following the precedent set for the Medicare FFS and Medicaid
analyses, this study defines a region as a Hospital Referral Region (HRR).
107
"CMS Manual System: Pub 100-20 One-Time Notification," Centers for Medicare & Medicaid Services,
http://www.cms.gov/Regulations-and-Guidance/Guidance/Transmittals/downloads/R978OTN.pdf.
108
In the FFS analysis, beneficiaries with a chronic condition are automatically enrolled in that chronic cohort in the
next year of the analysis period. Because health data for MA beneficiaries is only available yearly, however, MA
beneficiaries are only included in a chronic cohort if they have the relevant HCC for that year. The chronic cohorts
for the MA analysis include COPD, depression, and diabetes.
For both the aggregate and four condition-specific cohorts, this analysis applies three
exclusion restrictions to ensure that the estimated spending figures capture most (if not all) of a
beneficiarys healthcare spending. The MA approach broadly uses the same set of exclusion
criteria as the FFS analysis, although some of the institutional features of the MA data sources
require unique restrictions. Table F.3 presents the number of beneficiaries lost to each exclusion
restriction for the full 2007 through 2009 analysis. Because these exclusions need not be applied
in any order, some beneficiaries will be excluded based on more than one restriction. As a result,
the total percent of beneficiaries excluded is not necessarily equal to the sum of the beneficiaries
lost to each restriction.The analysis excludes beneficiaries whose enrollment data show that
Medicare was the secondary payer for their claims because Medicare is not paying the full cost
of these beneficiaries healthcare (column B in Table F.3). Beneficiaries who have Medicare as
the secondary payer are typically the working aged or working disabled. The analysis also
excludes beneficiaries with any negative net monthly CMS payment amount on their MA claims
because negative net payment amounts typically appear due to data errors from date mismatches
or double adjustments applied by CMS (column C). Finally, the analysis excludes beneficiaries if
their enrollment information cannot be matched to an entry in the EDB. In the aggregate cohort,
these restrictions remove 5.2 percent of Part C beneficiaries.
In addition to general enrollment restrictions, this analysis also applies a number of MA-
specific exclusion restrictions. Table F.4 presents the results of imposing these exclusion
restrictions at the monthly level for the 2007 through 2009 analysis. As the table describes, this
report removes observations for MA beneficiaries in a given month if:
This analysis excludes months beneficiaries are enrolled in demonstration planswhich CMS
uses to study the impact of potential MA program changesbecause these plans do not operate
under the same program rules as other MA plans (column B).109 Program of All-inclusive Care
for the Elderly (PACE) plans are managed care plans that are uniquely designed to provide
services to the frail elderly who would otherwise need nursing home level of care.110 Because the
minimum required services the plan covers are greater than the minimum required services for
other MA plans and the enrollment requirements are less restrictive, this analysis also excludes
months beneficiaries are enrolled in PACE plans (column C). Beneficiaries in MA plans (other
109
"Medicare Demonstration Projects & Evaluation Reports," Centers for Medicare & Medicaid Services,
http://www.cms.gov/Medicare/Demonstration-Projects/DemoProjectsEvalRpts/index.html.
110
"PACE Fact Sheet," Centers for Medicare & Medicaid Services, https://www.cms.gov/Medicare/Health-
Plans/pace/downloads/PACEFactSheet.pdf.
111
"Medicare Managed Care Manual, Chapter 8: CMS' Payments to Hospice Programs," Centers for Medicare &
Medicaid Services, http://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/mc86c08.pdf.
112
If a beneficiary has a missing or zero risk score in a given month, the analysis first assigns the earlier risk score in
the analysis period, if there is one available. If all earlier risk scores are missing or zero, the beneficiarys risk score
is built using HCCs.
The MA analysis calculates per capita monthly expenditures using three steps.
The initial step calculates expenditures for all beneficiaries enrolled in Part C. First, the
analysis sums all CMS capitated payments to MA and cost plans, any rebates the plans receive
from CMS for bidding below the benchmark, and all beneficiary premiums paid to the plans for
all months a beneficiary is enrolled in Part C. Although these premiums comprise the majority of
the spending for most MA beneficiaries, beneficiaries in cost plans may also receive FFS rates.
Thus, for cost plan beneficiaries, this analysis adds spending observed in the FFS claims. These
FFS payments are calculated as the sum of the claim payment, coinsurance, and deductible from
the Inpatient, Outpatient, Hospice, Home Health, Skilled Nursing, Carrier, and Durable Medical
Equipment claim types. For any MA beneficiary with prescription drug coverage through
Medicare, the analysis also adds all Part D spending during the months they are enrolled in Part
C. This analysis captures drug spending regardless of whether the beneficiary has prescription
drug coverage through a Medicare Advantage Prescription Drug (MA-PD) plan as part of their
MA plan or whether their coverage comes through a stand-alone prescription drug plan (PDP).
The expenditure level for a Part C beneficiary in any month is the sum of their MA premiums,
FFS payments (for cost plan beneficiaries), and prescription drug costs.
The second step adjusts these costs for inflation. As in the FFS analysis, the analysis uses
the Consumer Price Index to account for the changing prices paid by consumers for all goods and
services (not just health care).114 To inflate expenditures to costs in 2009, the analysis multiplies
expenditures incurred in 2007 by 1.032 and multiplies expenditures in 2008 by 0.995.
Finally, the third step calculates average expenditures per beneficiary per month for the
months they are enrolled in Part C. Typically, total beneficiary cost is the cost incurred during
113
Unlike FFS beneficiaries, the individual claims for each service Part C beneficiaries receive are not submitted to
CMS, except for prescription drug spending and services that beneficiaries in cost plans receive through FFS. Thus,
this analysis does not examine health care utilization counts or health care quality for Part C beneficiaries.
114
"Consumer Price Index - Chained Consumer Price Index".
This analysis examines price standardized cost to remove geographic variation in the
price of inputs. To create price-standardized figures, this methodology uses a four-step
approach. The study separately price-standardizes monthly payments to MA plans, payments for
FFS services received by beneficiaries in cost plans, and prescription drug spending. The
components are then renormalized so that the sum of price standardized expenditures equals the
sum of unadjusted expenditures.
First, monthly payments to MA plans and cost plans in a given region are price-
standardized based on the results of price standardization in that region for the Medicare FFS
analysis. Price standardized costs for region R are calculated by multiplying the unadjusted
expenditures for that region by a price standardization ratio using beneficiaries from the FFS
analysis in the aggregate cohort. The price standardization ratio is calculated as follows:
(F.1)
In equation (F.1), YSiR represents the price-standardized FFS cost for each beneficiary i in region
R for a given year, and YUiR represents the corresponding unadjusted cost for each beneficiary i in
region R.116 Cost is measured as allowed charges in Part A and Part B which includes payments
from Medicare as well as beneficiary coinsurance and deductible payments.117 Each regions
total standardized and unadjusted costs are calculated by summing costs for each beneficiary-
month of enrollment in FFS in that region. A PSR greater than one indicates that the regions
expenditures are relatively higher after accounting for input prices.
Second, the analysis price standardizes payments for beneficiaries in cost plans for
services they receive through FFS using the same price standardization methodology as the FFS
analysis. Price standardization occurs at the file type-year level for the Inpatient, Outpatient,
Hospice, Home Health, Skilled Nursing, Carrier, and Durable Medical Equipment claim types.
Third, payments for beneficiaries enrolled in Part C for prescription drugs are paid
through Part D, and this analysis price standardizes these costs using the same price
115
To assign beneficiaries to a region, this analysis utilizes the first ZIP code that appears during the beneficiarys
observation period.
116
Because this analysis price-standardizes costs on a monthly basis, beneficiaries costs are price-standardized
using the PSR of the region in which they currently reside.
117
In addition, because MA benchmarks include expected payments for indirect medical education and
disproportionate share hospitals, these costs have been included in calculating the FFS price standardization ratio.
In the fourth and final step, the total price standardized cost for all beneficiaries is
renormalized. Renormalization causes total price standardized costs to equal total unadjusted
costs. The analysis renormalizes price-standardized expenditures at the aggregate level; thus, for
the condition cohorts, total price standardized spending is not necessarily equal to total
unadjusted spending.
This analysis measures case mix based on the risk score CMS calculates as part of its
HCC model. The HCC model measures beneficiary demographics, enrollment status, and
severity of illness. CMSs HCC model uses separate models to calculate risk scores for different
categories of beneficiaries (e.g., community, institutional, and ESRD).118 As a result, beneficiary
risk scores represent health relative to other beneficiaries within that model. For example, a
beneficiary with ESRD may have a lower risk score than a beneficiary without ESRD, even
though the beneficiary with ESRD is likely to be relatively sicker and have higher health care
costs.
First, to standardize the risk score across models, this analysis creates a renormalized risk
score for each year. The renormalized risk score RNi for a beneficiary i in risk model M for a
given month is calculated as follows:
,
(F.2)
In equation (F.2), , represents the original risk score for beneficiary i in risk model M, and
is the average original risk score for all beneficiaries in risk model M. represents the
average monthly spending for beneficiaries in risk model M, while represents the average
118
This analysis utilizes risk scores after they are normalized by CMS to account for coding and population changes.
The normalized risk scores are used by CMS to determine payment.
a given risk model to 1.0, while the second term, , scales the risk scores by the average
expenditures for beneficiaries in that risk model as a proportion of total average expenditures
across all MA beneficiaries.
After renormalizing the risk scores, the second step separately risk adjusts MA premium
expenditures, cost plan FFS payments, and prescription drug expenditures, because beneficiary
health affects expenditures for these components in different ways. The risk adjustment approach
modifies capitated monthly MA payments by dividing total premiums paid for MA beneficiaries
by the renormalized risk score. CMS calculates payments to MA plans by multiplying the lower
of the unadjusted bid or the benchmark by beneficiary risk scores. By dividing the observed
premiums by the risk score, this approach, in essence, directly removes CMS adjustments for
health. Risk-adjusted spending to MA plans is then renormalized such that total risk-adjusted
spending is equal to total price-standardized spending, and risk-adjusted spending represents the
real prices paid by CMS and beneficiaries to MA plans.119
In the third and fourth steps, payments for beneficiaries in cost plans and prescription
drug spending are separately risk-adjusted using a regression-based approach. Unlike the MA
premiums, where the beneficiarys risk score directly affects CMS spending, beneficiary risk
scores affect cost plan and prescription drug spending only probabilistically (i.e., sicker
beneficiaries are more likely to have higher healthcare spending). To implement this approach,
the analysis estimates the following specification for each year of analysis:
(F.3)
In equation (F.3), represents the dependent variable for beneficiary i. represents the
coefficients estimating the relationship between the dependent and the independent variables
represented by Xi, and represents the error term. The two dependent variables that are
estimated using this method are average monthly price-standardized spending for beneficiaries in
cost plans and average monthly price-standardized spending on prescription drugs for
beneficiaries in Part C. This equation is estimated yearly at the beneficiary level using the OLS
regression method. 120
The outcome variable, Yi, is calculated on a monthly basis, or average spending per
month. Calculating average per-month costs removes the effect of increased expenditures at
119
Because the second and third steps use an OLS regression method, total risk adjusted spending is equal to total
price standardized spending.
120
To calculate risk-adjusted spending for beneficiaries in condition cohorts, each risk model is re-estimated on
those beneficiaries. In the 2007 through 2009 analysis, a year indicator is also included as an independent variable.
The residual for individual i, , is calculated as the difference between the observed (or
actual) value of the dependent variable, , minus the predicted value of the dependent
variable, where denotes the OLS estimates of the coefficients, . These risk
adjustment factors are added to average spending to create a risk adjusted value of spending for
each beneficiary in each year, , which can be expressed as follows:
(F.4)
In equation (F.4), is equal to the average predicted value of the dependent variable across all
beneficiaries in a given year. After determining the risk adjusted expenditures for all three
components, this analysis sums expenditures for each beneficiary to calculate total risk adjusted
spending per beneficiary.
This analysis risk-adjusts spending for beneficiaries in cost plans separately from
spending for other Part C beneficiaries because cost plans expenditures include FFS spending.
Thus, total average monthly expenditures for beneficiaries in cost plans are risk adjusted
together. The independent variable used to account for health status is the first renormalized risk
score RNi for beneficiary i during the observation period. By using the first renormalized risk
score of the observation period, this analysis imposes a prospective risk adjustment based on
beneficiary health status. This approach does not, however, account for any change in the
beneficiarys health status, including a change of risk model; for example, if a beneficiary moves
to the institutional setting, the resulting change in his or her risk model and risk score is not taken
into account.
Because prescription drug expenditures are paid at FFS rates rather than capitated rates
for beneficiaries in Part C, this analysis risk adjusts prescription drug spending separately from
both the monthly payments that MA plans receive and the costs for beneficiaries in cost plans.
The prescription drug spending risk adjustment includes all beneficiaries enrolled in Part C,
including those not enrolled in Part D. This analysis uses four independent variables to account
for patient health status and differences in enrollment. These specific variables include: (i) the
renormalized risk score (RNi); (ii) an indicator for continuous enrollment in Part D during the
entire time enrolled in Part C; (iii) a continuous Part D partial year enrollment variable, and (iv)
1. How much variation is there in per capita spending in Medicare Advantage across the
nation?
2. Are regions with high Medicare Advantage premiums likely to have high Medicare
Advantage premiums in subsequent years?
4. Are regions with high Medicare Advantage spending levels also likely to have high
spending levels in Medicare fee-for-service?
The remainder of this section addresses each of these questions in turn using the methodology
described in Sections F.2.1 through F.2.4.
121
In addition, 6 beneficiaries included the analysis did not have a sex listed in their enrollment files, likely due to
coding error.
Accounting for regional differences in input costs and beneficiary health characteristics
does reduce the variability in the MA spending distribution. Table F.7 shows the results of this
analysis. Because this analysis renormalizes risk adjusted MA spending for all beneficiaries, the
average monthly price-standardized risk adjusted cost is equal to the average unadjusted cost.123
The median cost, however, increases after risk adjustment from $719 to $937. Thus, the
skewness of the distribution decreases. Risk adjustment also decreases the average cost of death
episodes by almost half, although they remain the highest-cost episodes on average. Unlike the
FFS analysis, the relatively higher cost of episodes ending in death even after risk adjustment
does not indicate higher levels of health care utilization but instead mainly reflects the higher
premiums paid to MA plans for these beneficiaries. Further, risk adjustment decreases the
variation in average monthly costs; the standard deviation decreases by more than half, and the
90-10 difference decreases by two-thirds.
122
The exception to this rule is that utilization of healthcare services does affect spending for beneficiaries in Part C
cost plans or who are enrolled in a Part D prescription drugs plan.
123
This analysis does not, however, account for sex, race, dual-eligibility status, or death during the episode; thus,
the average risk adjusted spending is not equal for all beneficiary characteristics.
Although the MA risk adjusted cost distribution is right-skewed like the FFS cost
distribution, MA expenditures have less variation, are less skewed, and have higher averages
than FFS expenditures. Table F.8 displays the distribution of average monthly unadjusted and
price-standardized risk adjusted spending for MA beneficiaries and for FFS beneficiaries. First,
partly because monthly MA expenditures are based on benchmarks set by CMS, the distribution
of average monthly MA costs has less variability than the distribution of Medicare FFS costs; the
standard deviation of average monthly costs is $900 for MA and $2,032 for FFS. Though the
MA distribution is somewhat skewed, it is less skewed than the FFS distribution, which can be
seen by contrasting the median cost with the top and bottom 10th percentile. In the FFS
distribution, the lower tail, measured by the difference between the median and the 10th
percentile ($757), is just over half the size of the upper tail, measured by the difference between
the 90th percentile and the median ($1,360). In the MA distribution, however, the lower tail
($199) is about two-thirds the size of the upper tail ($298). Finally, average risk adjusted
spending is $986 for MA beneficiaries and only $958 for FFS beneficiaries. As a result of the
capitation update system and the budget neutrality adjustment, benchmarks for MA plans are set
above Medicares FFS costs; thus, aggregate payments for MA beneficiaries are usually higher
than they would be if the beneficiaries were enrolled in FFS, even after accounting for input
prices and beneficiary health status. The finding that average spending for MA beneficiaries is
higher than for FFS beneficiaries confirms MedPACs findings in their March 2012 Report to the
Congress.124
124
"The Medicare Advantage Program: Status Report".
Spending that depends on utilization, namely prescription drug spending, is not the
primary driver of regional variation in MA spending. Table F.9 presents the distribution of
unadjusted prescription drug spending alongside the distribution of total MA expenditures.
Prescription drug spending only accounts for 15 percent of total MA spending. The highest cost
MA beneficiaries, however, have a higher share of their total MA costs made up of Part D
spending. Whereas Part D spending makes up 11 percent of total spending for beneficiaries at
the median, Part D spending makes up 19 percent of total spending for beneficiaries at the 90th
percentile of total MA expenditures.125
125
Note that beneficiaries at the 90th percentile for Part D expenditures are not necessarily the same beneficiaries as
those at the 90th percentile for total MA expenditures.
Table F.12: Spearman Rank Correlation for Standardized Risk Adjusted MA Costs
2007 2008 2009
2007 1.000 0.950 0.910
Benchmark levels from 2007 through 2009 are even more stable. To determine the
stability of benchmarks at the HRR-level, Figure F.6 presents the 2007 benchmark for each HRR
against the 2008 benchmark, and Figure F.7 presents the 2008 benchmarks against the 2009
benchmarks. In 2008 CMS updated benchmarks. In years where CMS updates benchmarks,
each countys benchmark rate increases by the greater of two percent or the national per capita
MA growth percentage. Thus, county-level benchmarks are highly correlated from year to year,
and Figure F.6 illustrates that the 2007 and 2008 benchmarks are very closely related, as
indicated by the nearly perfect 45 degree line the bubbles form.126 The Pearson correlation
between 2007 and 2008 benchmarks is 0.999. In 2009, however, CMS rebased the
benchmarks. In rebasing years, CMS resets benchmark levels to the greater of the minimum
update or the per capita FFS spending in each county.127 As a result, all benchmarks increased
markedly from 2008 to 2009, as illustrated by the upward shift in Figure F.7. On average,
benchmarks increased by about 4 percent in real terms. In particular, the benchmark in Miami,
represented by the large bubble with the highest 2009 benchmark, increased 12 percent in 2009
because its FFS spending over recent years grew much faster than FFS spending at the national
level.128 Despite the fact that 2009 is a rebasing year, the Pearson correlation between 2008
and 2009 HRR-level benchmarks is 0.997.129 Thus, changes in each HRRs per capita MA
expenditure level are not due to changes in each HRRs benchmark rate.
126
After adjusting for inflation, benchmarks in 2008 are an average of 0.3 percent lower than benchmarks in 2007.
127
"The Medicare Advantage Program".
128
Though not discussed in this report, the Medicare FFS analysis found that after all claims processing completed
by June 2011, FFS expenditures in Miami actually increased 7.8 percent from 2008 to 2009. As noted by MedPAC
in their 2009 Report to the Congress cited above, Miami FFS expenditures in 2009 included millions of dollars in
payments for claims that have since been proved inappropriate.
129
After removing Miami as an outlier, the Pearson correlation between 2008 and 2009 benchmarks increases to
0.999.
(F.5)
In equation (F.5), niRt represents the number of beneficiaries in MA plan contract i in region R in
year t, and NRt represents the total number of beneficiaries in region R in year t. This analysis
calculates a single HHI for each region each year from 2007 through 2009 and then averages
these values to get a single 2007-2009 HHI value for each HRR. The MA plan HHI for a given
region ranges from zero to one. Higher HHI values indicate higher levels of market
concentration (i.e., lower levels of market competition). For instance, an HHI value of one
indicates that single plan has captured 100 percent of the HRRs market share.
Using an HRRs MA plan HHI, higher levels of market concentration have a small
negative correlation with average per capita MA expenditures. Figure F.8 presents a scatterplot
of average price-standardized, risk adjusted MA expenditures against the MA plan HHI. This
figure shows a weakly negative relationship between the HHI index and MA expenditures; that
is, less MA plan competition is associated with lower MA expenditures. The Pearson correlation
between the MA plan HHI and average monthly MA expenditures is -0.16.
Even after controlling for these market-level factors, HRRs with more concentrated MA
markets do not have higher spending levels. Table F.13 provides the coefficients and standard
errors (in parentheses) for each variable. In particular, the MA plan HHI has a negative
association with MA expenditures, indicating that even after controlling for other market-level
characteristics, more market concentration is associated with lower costs. This effect, however,
is not statistically significant and is relatively small in magnitude. Increasing an HRRs MA plan
HHI from 0.25 to 0.75 would decrease monthly per capita MA spending by about $25 or about
2.5 percent.
Variable Coefficient
-49.76
MA Plan HHI
(36.75)
-37.41
Hospital Beds HHI
(29.43)
26.23*
Teaching Hospital in HRR
(11.13)
10.7
Specialty Hospital in HRR
(13.2)
-10.94
Government-Owned Hospital in HRR
(9.63)
371.82***
Percent of Population Uninsured
(102.33)
446.47
Number of Hospitals per 1,000
(385.32)
14.36***
Number of Hospital Beds per 1,000
(4.19)
1.05*
Medicare Malpractice GPCI
(0.62)
13.84
Physicians per 1,000
(13.31)
Active Primary Care Physicians per -87.4*
1,000 (47.65)
Health Professional Shortage Area -8.96***
(HPSA) (2.15)
Average per capita MA expenditures are higher than average per capita FFS
expenditures ($986 and $958, respectively).
MA expenditures per person are less variable than FFS expenditures (with standard
deviations of $1,373 and $5,666, respectively).
Regions that are high- or low-cost in one year tend to be similarly high- or low-cost in
the next. The correlation between an HRRs per capita MA expenditures in 2007 and
its per capita MA expenditures in 2009 is 0.94.
HRRs with more MA plan competition do not necessarily have lower MA premiums.
In part because CMS uses a countys historical FFS spending to determine its
benchmark rate, the correlation between an HRRs per capita FFS and MA
expenditures is fairly high (0.66).
F.5 MA References