PGP Risk Report_2004_Model_Final

July 2006

PGP Demonstration Report on Risk
Adjustment
Final Report

Prepared for
John Pilotte
Centers for Medicare & Medicaid Services
7500 Security Boulevard
C4-17-27
Baltimore, MD 21244-1850

Prepared by
Eric Olmsted, Ph.D.
Gregory C. Pope, M.S.
John Kautter, Ph.D.
RTI International
411 Waverley Oaks Road, Suite 330
Waltham, MA 02452-8414

RTI Project Number 07964.013

PGP DEMONSTRATION REPORT ON RISK ADJUSTMENT

Authors:

Eric Olmsted, Ph.D.
Gregory C. Pope, M.S.
John Kautter, Ph.D.

Project Director:

Gregory C. Pope, M.S.

Federal Project Officer:

John Pilotte

RTI International*
CMS Contract No. 500-00-0024, T.O. No. 13
July 2006

This project was funded by the Centers for Medicare & Medicaid Services under Contract No.
500-00-0024, T.O. No. 13. The statements contained in this report are solely those of the authors
and do not necessarily reflect the views or policies of the Centers for Medicare & Medicaid
Services. RTI assumes responsibility for the accuracy and completeness of the information
contained in this report. The authors thank Michael Trisolini for his comments and suggestions.

*RTI International is a trade name of Research Triangle Institute.

CONTENTS
SECTION 1 INTRODUCTION AND OVERVIEW...................................................................... 1
1.1 Purpose of Risk Adjustment.......................................................................................... 1
1.2 Overview of the Risk Adjustment Process.................................................................... 1
1.3 Structure of Report ........................................................................................................ 2
SECTION 2 THE CMS-HCC RISK ADJUSTMENT MODEL..................................................... 3
2.1 Risk Marker Assignment............................................................................................... 3
2.1.1 Diagnostic Classification System......................................................................... 3
2.1.2 Hierarchies ........................................................................................................... 3
2.1.3 CMS-HCCs .......................................................................................................... 5
2.2 Expenditure Prediction.................................................................................................. 5
SECTION 3 RISK SCORES AND RISK ADJUSTMENT.......................................................... 11
3.1 Risk Scores.................................................................................................................. 11
3.2 Risk Adjustment of Expenditure Growth Rates and Medicare Savings
Calculations................................................................................................................. 12
3.2.1 Expenditure Growth Rates ................................................................................. 12
3.2.2. Medicare Savings ............................................................................................... 14
SECTION 4 CUSTOMIZATION OF THE CMS-HCC MODEL FOR PGP
DEMONSTRATION ........................................................................................................... 17
4.1 Concurrent versus Prospective Risk Adjustment ........................................................ 17
4.2 Recalibration of Model for PGP Demonstration Expenditures and Population.......... 19
4.3 ESRD Population ........................................................................................................ 19
4.4 Major Organ Transplants ............................................................................................ 19
4.5 New Enrollee Population ............................................................................................ 20
SECTION 5 PGP CONCURRENT RISK ADJUSTMENT MODEL.......................................... 21
5.1 Model Description....................................................................................................... 21
5.1.1 Model Variables ................................................................................................. 21
5.1.2 Sample Exclusions and Expenditures ................................................................ 22
5.1.3 Relative Weights ................................................................................................ 23
5.1.4 Constraints.......................................................................................................... 26
5.2 Demographic Adjustment ........................................................................................... 26
5.3 Risk Score Calculation ................................................................................................ 27
5.4 Summary ..................................................................................................................... 29
SECTION 6 NEW ENROLLEE AND ESRD MODELS............................................................. 31
6.1 PGP New Enrollee Model ........................................................................................... 31
6.1.1 Model Calibration and Variables ....................................................................... 31
6.1.2 Adjustment to Predict New Enrollee Mean Expenditures Accurately............... 32

i

6.1.3 Summary ............................................................................................................ 32
6.2 PGP ESRD Model....................................................................................................... 32
6.2.1 Defining ESRD Beneficiaries ............................................................................ 34
6.2.2 PGP ESRD Dialysis Model................................................................................ 36
6.2.3 Transplant Adjustment ....................................................................................... 40
6.2.4 Functioning Graft Adjustment ........................................................................... 42
6.2.5 PGP ESRD Model Risk Score Calculation ........................................................ 43
6.2.6 Summary ............................................................................................................ 44
SECTION 7 DATA REQUIREMENTS & MODEL UPDATES................................................. 45
7.1 Data Requirements ...................................................................................................... 45
7.2 Model Updates ............................................................................................................ 46
7.3 Upward Trend in Risk Scores ..................................................................................... 46
SECTION 8 CONCLUSION ........................................................................................................ 47
REFERENCES.............................................................................................................................. 49
List of Tables
Table 1
CMS Hierarchical Condition Categories.................................................................... 6
Table 3
Hypothetical Example of Risk Adjustment of Expenditure Growth........................ 13
Table 4
Hypothetical Example of Medicare Savings Calculation ........................................ 14
Table 5
PGP Concurrent Risk Adjustment Model for Continuing Enrollees Without
ESRD........................................................................................................................ 24
Table 6
PGP Concurrent Risk Adjustment Model Demographic Modifiers......................... 27
Table 7
Hypothetical Example of Initial Risk Score Calculation ......................................... 28
Table 8
PGP Demographic Model for New Enrollees Initial Risk Scores............................ 33
Table 9
PGP ESRD Dialysis Model...................................................................................... 38
Table 10
PGP ESRD Model—Transplant Relative Weights .................................................. 41
Table 11
PGP ESRD Model—Functioning Graft Adjustment ............................................... 43
Table 12
Hypothetical Example of ESRD Monthly Assignment............................................ 44
List of Figures
Figure 1
HCC Aggregations of ICD-9-CM Codes ................................................................... 4
Figure 2
Clinical Vignette for CMS-HCC Classification 79 Year Old Woman with AMI,
Angina Pectoris, COPD, Ankle Sprain, and Renal Failure........................................ 9
Figure 3
Application of the Overall Multiplier for the PGP New Enrollees Model............... 34
Figure 4
Hypothetical ESRD Status Assignment ................................................................... 43

ii

SECTION 1
INTRODUCTION AND OVERVIEW
The Physician Group Practice (PGP) demonstration is intended to reward improvements
in the efficiency of medical practice. The demonstration does this by creating a bonus pool based
on the growth in Medicare spending per beneficiary assigned to a PGP compared to the growth
in per beneficiary spending in a comparison group of beneficiaries. The rate of growth in per
beneficiary spending can also be affected by changes in casemix, or the health status, of the
beneficiaries in a group. As a result, the Benefits Improvement and Protection Act (BIPA)
legislation that authorizes the PGP demonstration requires that the performance targets be
adjusted for health “risk.”
1.1

Purpose of Risk Adjustment

To adjust for health risk, the PGP demonstration uses a version of the CMS-Hierarchical
Condition Category (HCC) model implemented for Medicare managed care risk adjustment. This
model, developed by RTI International under contract to CMS, is used to adjust capitation
payments to Medicare managed care Medicare Advantage (MA) plans (Pope et al., 2004). The
CMS-HCC model uses demographic information and diagnoses on administrative claims to
predict Medicare expenditures. The concurrent CMS-HCC model used in the PGP demonstration
is a modification of the prospective CMS-HCC model used to adjust managed care capitation
payments. The difference between the prospective and concurrent models is the prospective
model predicts expenditures from prior year diagnoses whereas the concurrent model predicts
expenditures from current year diagnoses. The reasons for this difference in risk adjustment
between the PGP demonstration and MA payment are discussed in Section 4.
The average risk score from the CMS-HCC model is applied to the observed per capita
expenditure growth rates to remove the effects of changes in health status. A PGP that treats a
population in the first year of the demonstration that is sicker than the population it treated in the
base year of the demonstration will have its per capita expenditure growth rate adjusted
downward to account for this change in health status. Similarly, the PGP’s expenditure growth
rate is adjusted upwards if the measured health status of its assigned population improves over
time. An example of these adjustments is presented in Section 3.
Risk adjustment in the PGP demonstration adjusts expenditure growth rates for changes
in average health status over time in PGP-assigned beneficiaries separately from the adjustments
made for comparison group beneficiaries. It is not an adjustment for differences at a point in time
between the health status of PGP-assigned and comparison group beneficiaries. For this reason,
even if casemix differs between PGP-assigned and comparison group beneficiaries, to the extent
that it is stable over time in these two groups, it will not affect comparison of PGP and
comparison group expenditure growth rates.
1.2

Overview of the Risk Adjustment Process

As previously mentioned, the model used in the PGP demonstration is a modification of
the CMS-HCC model currently used to adjust managed care capitation payments. The reasons

1

for this modification are discussed in Section 4, and are due to the role of risk adjustment in the
PGP demonstration.
Risk adjustment is used in the PGP demonstration according to the following steps:
•

assign risk markers;

•

predict expenditures;

•

calculate risk scores;

•

calculate population average risk scores;

•

adjust growth rates for risk;

•

calculate PGP expenditure target; and

•

compute Medicare savings.

Medicare savings are used to calculate the PGP bonus pool as described in the PGP
Demonstration Bonus Methodology Specifications report (Kautter et al., 2004). A PGP with a
lower adjusted growth rate than its comparison group generates Medicare savings and therefore a
bonus may be paid.
1.3

Structure of Report

This report describes the role of risk adjustment in the PGP demonstration, the
CMS-HCC model that is applied, and the adaptation of the CMS-HCC model for the PGP
demonstration. The next section describes how the CMS-HCC model uses diagnostic
information to predict expenditures for each beneficiary. Section 3 illustrates the calculation of
risk scores and how the risk scores are used to adjust expenditure growth rates. Section 4
provides an overview of how the CMS-HCC model has been adapted for the PGP demonstration.
Sections 5 and 6 describe the CMS-HCC models that will be used in the PGP demonstration.
Section 7 delineates the data requirements of the methodology, as well as how the model will be
updated over the life of the demonstration. Section 8 provides a brief conclusion.

2

SECTION 2
THE CMS-HCC RISK ADJUSTMENT MODEL
The CMS-HCC model is an expenditure prediction method based on health risk markers.
Risk markers are assigned using demographic and diagnostic information from health insurance
enrollment and claims files to create predictions of health care expenditures for Medicare
beneficiaries. These predictions are used to adjust per capita expenditure growth rates for any
changes that occur in the health status of the population under consideration.
This section describes the CMS-HCC model, including how risk markers are assigned
and used to generate health care expenditure predictions. The next section describes how health
care expenditure predictions are utilized in the PGP demonstration. The operation of the CMSHCC models and the model weights, are described in Sections 5 and 6.
2.1

Risk Marker Assignment
2.1.1 Diagnostic Classification System

The HCC diagnostic classification system begins by classifying each of the more than
15,000 ICD-9-CM diagnosis codes into 804 diagnostic groups, or DxGroups (see Figure 1).
Diagnosis codes are collected for each beneficiary over a twelve-month base period. Each ICD9-CM code maps into one DxGroup, which represents a specific medical condition. An example
is DxGroup 28.01 “acute liver disease”. DxGroups are further aggregated into 189 Condition
Categories, or CCs that describe major diseases and are broadly organized into body systems,
somewhat analogous to the ICD-9-CM major diagnostic categories. The CCs are designed to be
both clinically- and cost-similar, although they are not as uniform as the DxGroups. An example
is CC 28 “Acute Liver Failure/Disease” which includes DxGroups 28.01 and 28.02 “viral
hepatitis, acute or unspecified, with hepatic coma”. In most cases, DxGroups are assigned to only
one CC. However, in a few cases, a single ICD-9-CM code indicates more than one disorder, for
example, ICD-9-CM 404.03 “hypertensive heart and renal disease with congestive heart failure
and renal failure”. This code is assigned to DxGroup 131.03 “hypertensive heart/renal disease
with heart/renal failure”, which has a primary CC assignment of 131 “Renal Failure,” but also
receives a secondary or “duplicate” assignment to CC 80 “Congestive Heart Failure.”
2.1.2 Hierarchies
Hierarchies are imposed among related CCs so that a person is assigned only to one CC
with the most severe manifestation of related diseases. For example, ICD-9-CM ischemic heart
disease codes are organized into the “Coronary Artery Disease” hierarchy. The hierarchy
consists of 4 CCs arranged in descending order of clinical severity and cost, from CC 81 “Acute
Myocardial Infarction” to CC 84 “Coronary Athlerosclerosis/Other Chronic Ischemic Heart
Disease.” A person with an ICD-9-CM code for CC 81 is excluded from CCs 82, 83 or 84 even
if the person received ICD-9-CM codes that group for those categories. Similarly, a person with
ICD-9-CM codes that group into CCs 82 “Unstable Angina and Other Acute Ischemic Heart
Disease” and into 83 “Angina Pectoris/Old Myocardial Infarction” is assigned exclusively to CC
82.

3

Figure 1
HCC Aggregations of ICD-9-CM Codes

ICD-9-CM codes
(n = 15,000+)

DxGroups
(n = 804)

Condition Categories (CCs)
(n = 189)
Hierarchies
imposed
Hierarchical
condition categories (HCCs)
(n = 189)

CMS-Hierarchical
condition categories
(n = 70)

SOURCE: RTI International

After the hierarchies are imposed, the CCs become Hierarchical Condition Categories
(see Figure 1). In this way, the 15,000+ diagnosis codes are used to assign values for each
beneficiary for the full set of 189 HCCs. The value for each HCC can be either ‘1’ indicating that
the beneficiary has a diagnosis code for that condition, or ‘0’ indicating that the beneficiary does
not have a diagnosis code for that condition. A beneficiary can have multiple HCCs coded as ‘1’,
but not more than one in the same disease hierarchy.

4

2.1.3 CMS-HCCs
The CMS-HCC model selects only 70 of the original 189 HCCs for use in Medicare
Advantage payment. (Table 1 lists the 70 HCCs in the CMS-HCC model.) Thus, the CMS-HCC
model is a “selected significant diseases” model that focuses on adjusting for risk associated with
selected high-cost diagnoses; it does not incorporate all diagnoses. The 70 HCCs in the CMSHCC model:
•

cover a broad spectrum of health disorders;

•

contain well-defined diagnostic criteria;

•

include non-discretionary diagnoses in that they are serious disorders that are likely to
be diagnosed and treated when they occur; and

•

identify conditions with significant expected health expenditures.

HCCs that represent discretionary diagnoses that may or may not be diagnosed and/or
treated, and are subject to substantial diagnostic coding variations across providers were
excluded from the CMS-HCC system. Typically excluded HCCs are diseases or conditions with
a relatively low health and expenditure impact, such as HCC 24 “Other Endocrine/Metabolic/
Nutritional Disorders,” or vague or nonspecific HCCs such as, HCC 167 “Minor Symptoms,
Signs, Findings.” Excluded HCCs also include diseases that are highly prevalent among
Medicare beneficiaries but subject to erratic diagnosis and coding such as HCC 91
“Hypertension.”
In addition to diagnosis based markers, the CMS-HCC model uses a variety of
demographic markers. Demographic markers are based on the age, sex, and enrollment status of
the beneficiary. The enrollment status includes whether the beneficiary is enrolled in Medicaid,
or was originally qualified for Medicare due to disability. Medicare beneficiaries under 65 years
of age qualify because of disability.
The CMS-HCC model greatly reduces administrative complexity while sacrificing little
predictive power compared to the full 189 HCC model. Beneficiaries diagnosed with at least one
CMS-HCC encompass 61% of all Medicare fee-for-service beneficiaries, but they account for
94% of total expenditures for all Medicare fee-for-service beneficiaries.1 Also, the CMS-HCC
model explains 92% of the variation in health care expenditures that is explained by including all
189 HCCs. The CMS-HCC model creates predictions that are more robust to diagnostic coding
and treatment differences across providers than the full model.
2.2

Expenditure Prediction

Risk markers are the building blocks with which health care expenditure prediction is
based. Each of the risk markers (including both HCCs and demographic markers) in the CMS-

1

Costs of beneficiaries without any CMS-HCCs are predicted with demographic information (costs of
beneficiaries with at least one CMS-HCC are predicted with both diagnostic and demographic information).
Thus, expenditures are predicted for all beneficiaries and all costs are included in the model.

5

Table 1
CMS Hierarchical Condition Categories
HCC Number

HCC Label

HCC1
HCC2
HCC5
HCC7
HCC8
HCC9
HCC10
HCC15
HCC16
HCC17
HCC18
HCC19
HCC21
HCC25
HCC26
HCC27
HCC31
HCC32
HCC33
HCC37
HCC38
HCC44
HCC45
HCC51
HCC52
HCC54
HCC55
HCC67
HCC68
HCC69
HCC70
HCC71
HCC72
HCC73
HCC74
HCC75
HCC77
HCC78
HCC79
HCC80

HIV/AIDS
Septicemia/Shock
Opportunistic Infections
Metastatic Cancer and Acute Leukemia
Lung, Upper Digestive Tract, and Other Severe Cancers
Lymphatic, Head and Neck, Brain, and Other Major Cancers
Breast, Prostate, Colorectal and Other Cancers and Tumors
Diabetes with Renal or Peripheral Circulatory Manifestation
Diabetes with Neurologic or Other Specified Manifestation
Diabetes with Acute Complications
Diabetes with Ophthalmologic or Unspecified Manifestation
Diabetes without Complication
Protein-Calorie Malnutrition
End-Stage Liver Disease
Cirrhosis of Liver
Chronic Hepatitis
Intestinal Obstruction/Perforation
Pancreatic Disease
Inflammatory Bowel Disease
Bone/Joint/Muscle Infections/Necrosis
Rheumatoid Arthritis and Inflammatory Connective Tissue Disease
Severe Hematological Disorders
Disorders of Immunity
Drug/Alcohol Psychosis
Drug/Alcohol Dependence
Schizophrenia
Major Depressive, Bipolar, and Paranoid Disorders
Quadriplegia, Other Extensive Paralysis
Paraplegia
Spinal Cord Disorders/Injuries
Muscular Dystrophy
Polyneuropathy
Multiple Sclerosis
Parkinson's and Huntington's Diseases
Seizure Disorders and Convulsions
Coma, Brain Compression/Anoxic Damage
Respirator Dependence/Tracheostomy Status
Respiratory Arrest
Cardio-Respiratory Failure and Shock
Congestive Heart Failure
(continued)

6

Table 1 (continued)
CMS Hierarchical Condition Categories
HCC Number

HCC Label

HCC81
HCC82
HCC83
HCC92
HCC95
HCC96
HCC100
HCC101
HCC104
HCC105
HCC107
HCC108
HCC111
HCC112
HCC119
HCC130
HCC131
HCC132
HCC148
HCC149
HCC150
HCC154
HCC155
HCC157
HCC158
HCC161
HCC164
HCC173
HCC174
HCC176
HCC177

Acute Myocardial Infarction
Unstable Angina and Other Acute Ischemic Heart Disease
Angina Pectoris/Old Myocardial Infarction
Specified Heart Arrhythmias
Cerebral Hemorrhage
Ischemic or Unspecified Stroke
Hemiplegia/Hemiparesis
Cerebral Palsy and Other Paralytic Syndromes
Vascular Disease with Complications
Vascular Disease
Cystic Fibrosis
Chronic Obstructive Pulmonary Disease
Aspiration and Specified Bacterial Pneumonias
Pneumococcal Pneumonia, Emphysema, Lung Abscess
Proliferative Diabetic Retinopathy and Vitreous Hemorrhage
Dialysis Status
Renal Failure
Nephritis
Decubitus Ulcer of Skin
Chronic Ulcer of Skin, Except Decubitus
Extensive Third-Degree Burns
Severe Head Injury
Major Head Injury
Vertebral Fractures without Spinal Cord Injury
Hip Fracture/Dislocation
Traumatic Amputation
Major Complications of Medical Care and Trauma
Major Organ Transplant Status (Procedure)
Major Organ Transplant Status
Artificial Openings for Feeding or Elimination
Amputation Status, Lower Limb/Amputation Complications

SOURCE: RTI International

7

HCC model is assigned a dollar value based on its predicted impact on health care expenditures.
A prediction is generated for every beneficiary by summing the dollar amounts for the
corresponding HCC and demographic markers assigned to the beneficiary. The total is the
beneficiary’s predicted health care expenditure for the analysis year.2
As an example of expenditure prediction, consider our hypothetical scenario in Figure 2
of a 79-year-old woman diagnosed with AMI, angina pectoris, COPD, renal failure, and an ankle
sprain over a twelve month period. The seven reported diagnosis codes assign five HCCs which
are used to create an expenditure prediction. The woman receives the incremental cost
predictions from a preliminary version of the concurrent CMS-HCC model shown in Table 2.
Note that not every diagnosis is used to generate the expenditure prediction. The
CMS-HCC model is a hierarchical model, and the woman receives no incremental cost
prediction for angina pectoris because AMI is ranked higher in the coronary artery disease
hierarchy. No incremental prediction is made for ankle sprain because this diagnosis is not
included in the CMS-HCC model. Ankle sprain is an example of a condition excluded because it
has a relatively low impact on expenditure, and it may not always be diagnosed or treated. Her
total expenditure prediction is the sum of the incremental predictions, or $21,870.

2

For comparison of beneficiary groups, expenditure predictions are converted to risk scores. This process is
described in Section 3. The models presented in Sections 5 and 6 present risk score coefficients, rather than
dollar coefficients.

8

Figure 2
Clinical Vignette for CMS-HCC Classification
79 Year Old Woman with AMI, Angina Pectoris, COPD, Ankle Sprain, and Renal Failure

1. DxGroup - Diagnosis Group
2. CC - Condition Category
3. HCC - Hierarchical Condition Category
4. HCC 83 is superceded by HCC 81 within the coronary disease hierarchy. HCC 81 is the more severe
manifestation and is, therefore, included.
SOURCE: RTI International

9

Table 2
Hypothetical Example of Expenditure Prediction
Risk Marker

Incremental Prediction

AMI (HCC 81)

$14,629

Angina pectoris (HCC 83)1

$0

COPD (HCC 108)

$2,465

Renal failure (HCC 131)

$4,776

Ankle sprain (HCC 162)2

$0

TOTAL

$21,870

1

HCC 83, angina pectoris has an incremental prediction, but the amount is not added
because HCC 81, AMI, is within the same hierarchy and is the more severe manifestation
of cardiovascular disease.

2

Ankle sprain is excluded due to its low impact on expenditures.

SOURCE: RTI International.

10

SECTION 3
RISK SCORES AND RISK ADJUSTMENT
Risk scores are comparisons of predicted expenditures for a beneficiary to the average
expenditures of all Medicare beneficiaries. This section describes the calculation of risk scores,
and discusses how risk scores are used to adjust expenditures and the calculation of Medicare
savings. A hypothetical example is included to clarify the concepts and methodology.
3.1

Risk Scores

Each beneficiary in a sample population generates a total expenditure prediction based on
the risk markers assigned, and each expenditure prediction is used to calculate an individual’s
risk score. Individual risk scores are then used to calculate average risk scores for the entire
population.
The risk score is the ratio of the beneficiary’s predicted expenditure and the average
expenditures of all Medicare beneficiaries. The risk score expresses how expensive a beneficiary
is predicted to be relative to the “average” Medicare beneficiary.
Risk Score =

Beneficiary’s Predicted Expenditure
National Average of Medicare
Beneficiaries’ Expenditures

The national average expenditure for the Medicare population was $7,7283 in 2004 (the
model is calibrated based on the experience of beneficiaries in the year 2004). Therefore, a
beneficiary who has predicted expenditures of $7,728 will have a risk score of 1.000. A
beneficiary who has predicted expenditures of $15,456 will have a risk score of 2.000, and has
double the expenditure risk of the average Medicare beneficiary.
The average of risk scores for individual beneficiaries weighted by person years of
eligibility generates the average risk score for the population under consideration. A PGP that is
assigned 15,000 full-year-eligible beneficiaries has an average risk score equal to the sum of the
15,000 individual risk scores divided by 15,000.
Average Risk Score4 =

3

4

Sum of Beneficiary Risk Scores for Group
Number of Beneficiaries in Group

The actual national average expenditure was calibrated using the PGP demonstration model calibration sample.
The PGP sample includes only beneficiaries with at least one E&M visit during 2004. See Section 4 for a
discussion of the PGP sample.
This formula assumes that all beneficiaries have 12 months of enrollment. Actual average risk score calculations
will use the sum of the fraction of months enrolled (i.e., full-year equivalents) for the beneficiaries as the
denominator.

11

Population risk scores are interpreted similarly to individual risk scores. A population
with a risk score greater than 1.000 indicates expected expenditures greater than average. A
population with a risk score less than 1.000 indicates expected expenditures less than average.
3.2

Risk Adjustment of Expenditure Growth Rates and Medicare Savings Calculations
3.2.1 Expenditure Growth Rates

The average risk score for a performance year is compared to the average risk score for
the base year5 to create risk ratios, which are then used to adjust base year per capita
expenditures. The risk ratio is created by dividing the average risk score for the population
during the performance year by the population average risk score during the base year. Risk
ratios are created separately for each PGP and each PGP’s comparison group.
Risk Ratio =

Average Risk Score in Performance Year
Average Risk Score in Base Year

A PGP’s or comparison group’s risk ratio adjusts the observed base year per capita
expenditures which is then compared to the performance year per capita expenditures to
calculate the risk adjusted growth rate. A PGP that is assigned a set of beneficiaries with a higher
average risk score in the performance year than in the base year will have its base year
expenditures adjusted higher, reducing the adjusted growth rate.

Adjusted Base Year Per Capita Expenditures =

Base Year Per Capita
Expenditures * Risk Ratio

Adjusted base year per capita expenditures are calculated for both the PGP and
comparison group beneficiary populations. Adjusted comparison group per capita growth rates
set the performance target for the PGP, and are used to evaluate PGP efficiency for that year.
Adjusted Per Capita Growth Rate =
(Actual Performance Year Per Capita Expenditures-Adjusted Base Year Per Capita Expenditures)
Adjusted Base Year Per Capita Expenditures
The performance target for the PGP is equal to the adjusted per capita growth rate for the
comparison group multiplied by the PGP’s adjusted base year per capita expenditures. The
difference between the PGP target per capita expenditures and actual per capita expenditures
generates the bonus pool for the participating PGP.

5

The base year for the PGP demonstration will be April 2004 to March 2005.

12

Adjusted Comparison Group Per Capita
Growth Rate * Adjusted PGP Base Year
Per Capita Expenditures

PGP Performance Target =

The example provided in Table 3 illustrates the importance of adjusting for health risk
when comparing expenditure growth rates. The first row shows the observed expenditures and
risk scores of a PGP during a demonstration performance year. Per capita expenditures have
grown from $6,000 in the base year to $6,400 in the performance year, for an unadjusted growth
rate of 6.7%.6 The average risk score of the assigned beneficiaries has also risen, from 1.00 to
1.05, indicating that the average health status of the beneficiaries assigned to the PGP has
declined. The risk ratio7 is applied to the base year expenditures to adjust for the change in the
health status of the beneficiaries assigned to the PGP. The adjusted base year expenditures are
$6,3008 resulting in a risk adjusted growth rate of only 1.6%.9 The health status of beneficiaries
assigned to the PGP in the performance year compared to the base year was such that per capita
expenditures are expected to grow from $6,000 to $6,300 due only to differences in health status.
Table 3
Hypothetical Example of Risk Adjustment of Expenditure Growth
Actual Per Capita
Expenditures
Base
Year

Performance
Year

Growth
Rate

Average Risk Score
Base Performance
Year
Year

Risk
Ratio

Risk Adjusted
Expenditures, Expenditures, Growth
Base
Performance
Rate

PGP
Beneficiaries 6,000

6,400

6.7%

1.000

1.050

1.05

6,300

6,400

1.6%

Comparison
Group

6,630

2.0%

1.000

0.950

0.95

6,175

6,630

7.4%

6,500

The expenditures of beneficiaries in the comparison group grew at only a 2.0%10
unadjusted rate. In contrast to the beneficiaries assigned to the PGP, the average risk score of the
comparison group beneficiaries has declined from 1.000 to 0.950, indicating an improvement in
the health status of those beneficiaries. As a result the comparison group risk ratio equals 0.95.11

6

(6,400 – 6,000 ) / 6,000 = 6.7%.

7

1.050 / 1.000 = 1.05.

8

6,000 * 1.050 = 6,300.

9

(6,400 – 6,300) / 6,300 = 1.6%.

10

(6,630 – 6,500) / 6,500 = 2.0%.

11

.950 / 1.000 = .95

13

This is applied to the comparison group base year expenditures, resulting in adjusted base year
expenditures of $6,175.12 The resulting risk adjusted comparison group growth rate is 7.4%.13
3.2.2. Medicare Savings
Medicare savings, which comprise the potential bonus pool, are calculated by comparing
actual performance year expenditures to the PGP’s target expenditures. Target expenditures are
equal to the PGP’s base year expenditures multiplied by the comparison group’s growth rate.
Consider the example in Table 4 that shows this calculation with and without risk adjustment.
The top row shows the calculation of per capita PGP target expenditures and Medicare savings
without using risk adjustment. In this case, target expenditures would be simply equal to actual
base year expenditures multiplied by the comparison group’s actual expenditure growth rate, or
$6,120.14 Per capita Medicare savings without risk adjustment would be -$280, the difference
between target expenditures and actual PGP performance year expenditures.15
Table 4
Hypothetical Example of Medicare Savings Calculation
Per Capita

Unadjusted
Risk Adjusted

PGP Base
Year
Expenditures

Comparison
Group
Expenditure
Growth Rate

6,000
6,300

2.0%
7.4%

Actual PGP
Performance
PGP Target
Year
Medicare
Expenditures Expenditures Savings
6,120
6,766

6,400
6,400

–280
366

Risk adjusted expenditures provide a more accurate assessment of the performance of the
PGP. Notice that PGP base year expenditures are 6,300 in the second (risk adjusted) row. As
previously mentioned, this indicates that the change in health status from the base year to the
performance year of PGP assigned beneficiaries would have driven per capita expenditures up
$300. Risk adjusted target expenditures equal to $6,766 are now calculated as risk adjusted PGP
base year expenditures multiplied by the risk adjusted comparison group growth rate.16 Actual
per capita Medicare savings are therefore $366, the difference between risk adjusted PGP target
expenditures and actual PGP performance year expenditures.17
12

6,500 * .95 = 6,175.

13

(6,630 – 6,175) / 6,175 = 7.4%.

14

6,000 * 1.020 = 6,120.

15

6,120 – 6,400 = -280.

16

17

6,300 * 1.074 = 6,766.
6,766 – 6,400 = 366.

14

The above method is used to calculate target expenditures and Medicare savings during
each performance year of the demonstration. Comparison group expenditure growth rates are
measured from the same base year (i.e., April 2004 to March 2005) for each Performance Year 1,
2, and 3. PGP performance in each Performance Year, therefore, depends on cumulative
expenditure growth since the base year. The demonstration is not rebased during its three year
duration. The base year for the demonstration is always used to calculate expenditure targets.
In the example presented in Table 4, unadjusted expenditures of PGP-assigned
beneficiaries grew at a higher rate than expenditures of the comparison group. When adjusted for
health risk the relative growth rates are reversed and the PGP may be eligible for a bonus, as
seen by the positive Medicare savings. Conversely, in some cases, risk adjustment could also
eliminate a PGP’s eligibility for bonuses calculated using unadjusted data.

15

[This page left intentionally blank.]

16

SECTION 4
CUSTOMIZATION OF THE CMS-HCC MODEL FOR PGP DEMONSTRATION
The primary modification made to the prospective CMS-HCC model for the PGP
demonstration was to develop a concurrent version of the model (concurrent models are
discussed in Section 4.1). In addition, we refined the concurrent model to meet the needs of the
PGP demonstration. These modifications fall into the following categories:
•

recalibrating the model to reflect the expenditures and population eligible for the PGP
demonstration;

•

including beneficiaries entitled by end-stage renal disease (ESRD);

•

identifying beneficiaries receiving a major organ transplant in a performance year;
and

•

including beneficiaries who are newly enrolled in Medicare during a performance
year.

The PGP demonstration uses the same 70 HCCs in its concurrent model as are used by
CMS for the prospective payment model, but additional risk markers are used to account for
beneficiaries entitled by ESRD, and those that received a major organ transplant. ESRD
beneficiaries will be included in the PGP demonstration so they must be accounted for in the
concurrent risk adjustment model. Beneficiaries who have received organ transplants are
included in the prospective CMS-HCC model calibration sample, but are not explicitly identified
by a transplant procedure code in the year of their transplant. Transplant recipients are very
expensive in the year they receive their transplant, so it is important to adjust for them in
concurrent risk adjustment. In addition, a separate methodology is used to calculate predictions
for new enrollees. Finally, the model is recalibrated to reflect expenditures of the beneficiary
population eligible for the PGP demonstration. In particular, the recalibration sample is restricted
to users of office or other outpatient evaluation and management (E&M) services.
The cost patterns of these groups need to be explicitly recognized in the concurrent risk
adjustment model. The next section explains the need for the concurrent model in the PGP
demonstration, while the succeeding sections describe the steps taken to adjust the concurrent
CMS-HCC model for beneficiaries eligible for the demonstration.
4.1

Concurrent versus Prospective Risk Adjustment

The CMS-HCC model used for Medicare Advantage plan payment is “prospective” in
that it uses prior year diagnoses to predict Medicare expenditures. The “concurrent” model
applied in the PGP demonstration uses current year diagnoses to predict Medicare expenditures.
A prospective risk adjustment model places more emphasis on chronic conditions that are likely
to affect health care costs during future periods. This is preferable when making capitation
payments in advance. In contrast, concurrent models capture acute illnesses (including acute
exacerbations of chronic illnesses) that have higher costs during the performance year.

17

A beneficiary that experiences an acute event, such as a heart attack, is expected to have
somewhat elevated expenditures in the following year, but will have significantly elevated
expenditures during the year the heart attack occurs. The prospective model puts an MA plan at
risk for the occurrence of the heart attack in a particular person. It compensates on average with
payments for such events through dollars associated with the demographic profile and with
chronic conditions associated with higher risk of heart attack in the following year. A concurrent
model accounts for the higher current expenditures of current-year heart attack patients.
There are several reasons to use concurrent rather than prospective risk adjustment in the
PGP demonstration. First, the PGP demonstration is a non-enrollment model, with assignment of
beneficiaries to PGPs based on current-year utilization. Only concurrent risk adjustment can
account for the non-random assignment of beneficiaries to PGPs based on current year health
status. Consider triaging referral of acute care cases. Some PGPs participating in the
demonstration may be tertiary care referral centers. The most serious, complex cases would be
referred to them based on acute, emergent conditions. The health status and expenditure risk
posed by these cases can be measured only by concurrent risk adjustment utilizing current
diagnoses. Prospective risk adjustment using last year's diagnoses cannot measure emergent
acuity and would be inadequate for the PGP demonstration.
Prospective risk adjustment is appropriate for MA risk adjustment because beneficiaries
must enroll in MA plans, which are then responsible for all their care over a period of time.
Assignment of beneficiaries to MA plans occurs at the beginning of the period (typically the
beginning of the year), and is not changed based on emergent variations in health status. Thus, it
is appropriate to adjust the risk of MA plans based on information known at the time of
enrollment, which is the information used in prospective risk adjustment.
Second, concurrent models explain a much higher proportion of expenditure variation
than do prospective models. The percentage of individual variation explained by a concurrent
model is approximately 50%, versus approximately 10% for prospective models (Pope et al.,
2000). This makes concurrent models more accurate in adjusting expenditure growth rates for
health status. The reason for the higher explanatory power of concurrent models is that they
explain expenditure variations associated with acute events in the current year that prospective
models will miss. This means concurrent models greatly reduce performance risk related to
health status variation compared with prospective models18.
HMOs and other MA organizations are licensed risk-bearing entities that can assume the
risk related to prospective risk adjustment. In contrast, PGPs participating in the PGP
demonstration are provider groups that are not at risk in the demonstration, although they have
an opportunity to earn a bonus. What is needed for the PGP demonstration is a “casemix”
adjuster to control for the mix of cases actually seen in the present year, not a prospective “risk”
adjuster to control for future risk based on prior information.19
18

Concurrent models also give credit for complications that occur during the current year.

19

The terms “casemix” adjustment and “risk” adjustment are often used somewhat loosely and interchangeably. A
more consistent usage of the terms would associate “casemix” adjustment with concurrent risk adjustment for
present time periods, and “risk” with prospective risk adjustment for future time periods.

18

Third, in practice, MA capitation rates are set at the beginning of the year. MA plans
typically want to know what “budget” they have to manage within. Setting rates at the beginning
of the year requires using information available at that time, which is the prior year diagnoses
used in prospective risk adjustment. On the other hand, in the PGP demonstration, bonus
calculations will occur retrospectively, after the end of each performance year when complete
claims data are available. This retrospective time frame makes concurrent risk adjustment
feasible for the PGP demonstration.
4.2

Recalibration of Model for PGP Demonstration Expenditures and Population

There are some differences between the expenditures and beneficiary population eligible
for the PGP demonstration versus the expenditures and sample used in estimating risk
adjustment models for Medicare managed care. Most importantly, in the PGP demonstration,
annualized per beneficiary expenditures are capped at $100,000 and beneficiaries must have at
least one office or other outpatient evaluation and management service to be eligible for the
demonstration (PGP-assigned or comparison group) (Kautter et al., 2004). In managed care risk
adjustment modeling, expenditures are not capped and the sample is not restricted according to
beneficiary utilization. These differences can affect measured health risk and expenditure
predictions. To account for the differences, RTI recalibrated the concurrent CMS-HCC model
using the expenditure definition and sample eligible for the PGP demonstration. In the next
section we present the model produced after modification and recalibration.
4.3

ESRD Population

Approximately 1% of Medicare beneficiaries are entitled by ESRD. Although this is a
small proportion, ESRD eligibles are, on average, nearly 10 times more expensive than
beneficiaries entitled by age or disability (an average annualized cost of approximately $60,000
for ESRD beneficiaries compared to close to $7,000 for aged/disabled beneficiaries). To account
for these cost differences the concurrent PGP demonstration model was adjusted to capture the
mean costs of:
•

ESRD enrollees currently undergoing dialysis;

•

ESRD enrollees undergoing a kidney transplant; and

•

ESRD enrollees that have already had a kidney transplant (and are maintaining a
functioning graft).

A separate prospective risk adjustment model has been developed by CMS for capitated
MA payment for beneficiaries entitled by ESRD. We use a similar, though simpler version of the
model for the PGP demonstration. This approach accounts for the high average and concurrent
costs of ESRD beneficiaries and for their diagnostic profile.
4.4

Major Organ Transplants

Beneficiaries who receive a major organ transplant (bone marrow, heart, liver, lung,
pancreas, intestines) are also substantially more expensive than an average Medicare beneficiary.
The concurrent CMS-HCC model used for the PGP demonstration includes a HCC risk category
19

for major organ transplants to capture their very high current year expenditures. This category is
based on CPT procedure codes recorded on claims, unlike the ICD-9 diagnosis codes used for
most HCCs.
4.5

New Enrollee Population

We developed a demographic model to predict expenditures for new enrollees. The PGP
demonstration requires that eligible beneficiaries have Part A and Part B coverage for all of the
months they are enrolled in Medicare during a demonstration year. We therefore define new
enrollees as beneficiaries eligible for the demonstration who are not continuously enrolled in
both Part A and Part B Medicare for all of their months alive during a demonstration year. (New
enrollees must have at least one month of A/B enrollment—and no months of A-only or B-only
enrollment—during a demonstration year to be eligible for the demonstration.) A beneficiary is
considered continuously enrolled:
•

if they were enrolled in January of the demonstration year; and

•

if their Part A and Part B coverage is continuous through December of that year, or
until the death of the beneficiary.

All other beneficiaries eligible for the demonstration are considered new enrollees. For
example, a beneficiary newly enrolling in the Medicare program at 65 years of age in the middle
of a demonstration performance year is considered a new enrollee. Continuing enrollees are risk
adjusted using the CMS-HCC model, new enrollees are not. Diagnosis-based risk adjustment
requires a complete diagnostic profile, which is not available for new enrollees. New enrollees,
therefore, receive an expenditure prediction from the Medicare Advantage (MA) demographic
model, which has been recalibrated for the PGP demonstration population.20 This model is
currently used for risk adjustment of aged or disabled beneficiaries enrolling in MA plans for
which the CMS-HCC model is inapplicable. The PGP demonstration model will apply a
prediction based solely on the age, sex, and Medicaid status of the beneficiary, weighted for the
number of months that the beneficiary was enrolled in both Part A and Part B Medicare.

20

See Section 6.1, “PGP New Enrollee Model”.

20

SECTION 5
PGP CONCURRENT RISK ADJUSTMENT MODEL
This section describes the risk adjustment model used in the PGP demonstration for
beneficiaries continuously enrolled in Medicare for an entire performance year (or until their
death), and who do not have end stage renal disease (ESRD). We describe the model and its
calibration and provide an example of risk score calculation.
5.1

Model Description

This section presents the PGP concurrent risk adjustment model for aged/disabled
continuing enrollees without ESRD. This model is used to create risk scores for beneficiaries that
are continuously enrolled in Medicare for the entire performance year (or until date of death),
and are not identified as ESRD beneficiaries (risk adjustment for new Medicare enrollees and for
ESRD beneficiaries is described in Section 6). Beneficiaries enrolled in Medicare because of age
or disability at the beginning of a base or demonstration year will be given a risk score from this
model. The PGP concurrent model incorporates the CMS-HCC risk markers described in
Section 2.
Creating risk scores using the PGP concurrent model follows the four-step process below:
1. Assign risk markers and demographic category.
2. Attach relative weights.
3. Calculate initial risk score.
4. Modify risk score for demographic category.
5.1.1 Model Variables
The PGP concurrent model is built from the prospective CMS-HCC model used by
Medicare to pay MA plans. Whereas the prospective CMS-HCC model uses CMS-HCCs based
on last year's diagnoses to predict this year's expenditures, the PGP concurrent model uses CMSHCCs based on this year’s diagnoses to predict this year's expenditures. The CMS-HCC model
uses 70 of the 189 hierarchical condition categories (HCCs). The 70 CMS-HCCs were selected
based on the clinical expectation of beneficiaries with these conditions incurring significant
medical expenditures. A list of these CMS-HCCs is provided as Table 1 in Section 2.
To reduce administrative burden, the prospective CMS-HCC model used to set payment
rates for MA plans requires plans to report only diagnosis codes (not procedures). However, the
PGP demonstration has access to FFS claims and therefore to procedure codes. The model
developed for the PGP demonstration takes advantage of this by including an HCC whose
assignment is based on transplant procedure codes found in the claims data.21 Lung, heart, liver,
21

As described in Section 4.

21

bone marrow, intestine, and pancreas transplants are indicative of very high expenditures and are
therefore included in the model as HCC 173 Major Organ Transplant (Procedure). This results in
a total of 71 CMS-HCCs that are included in the PGP concurrent model.22
In addition to the CMS-HCCs included in the model, we added a variable that indicates a
beneficiary has none of the 71 CMS-HCCs, which we call the “NOCMSHCC” variable.
Beneficiaries with at least one of the CMS-HCCs account for more than 90% of all Medicare
expenditures, but beneficiaries without any of these diseases may be diagnosed with other
conditions. These beneficiaries will utilize medical services, and therefore generate expenditures.
The NOCMSHCC variable provides a constant prediction of beneficiary costs for those
beneficiaries that have none of the significant diseases incorporated in the CMS-HCC model, but
nevertheless incur medical costs during the year. The NOCMSHCC variable is only assigned to
beneficiaries that do not have any of the 71 CMS-HCCs, and therefore represents the average
cost for a beneficiary identified as having none of those conditions. Beneficiaries in this category
therefore all receive the same relative weight.23
In summary, the PGP concurrent risk adjustment model uses 71 CMS-HCCs, as well as
the NOCMSHCC variable, to predict expenditures and generate risk scores.
5.1.2 Sample Exclusions and Expenditures
To develop the PGP concurrent model, RTI analyzed Medicare claims from the year
2000 for a 5% national random sample of FFS Medicare beneficiaries. We restricted the
estimation sample to beneficiaries with characteristics of those who will be eligible for the PGP
assigned or comparison group beneficiaries in the demonstration. To mimic the specifications of
the PGP demonstration, we applied the sample exclusions listed below (sample selection of new
Medicare enrollees and ESRD beneficiaries eligible for the demonstration is discussed in Section
6).
To be eligible for the 2004 PGP concurrent risk adjustment model calibration sample, a
beneficiary must:

22

23

•

be alive and enrolled in Medicare on January 1, 2004;

•

have a record in the Medicare enrollment file;

•

be enrolled in both Part A and Part B for all months of Medicare enrollment during
2004;

•

have at least one month of fee-for-service, aged/disabled, non-hospice Medicare
enrollment in 2004;

The PGP concurrent risk adjustment model includes the 70 CMS-HCCs from the Medicare Advantage (MA) risk
adjustment model, plus HCC 173 Major Organ Transplant (Procedure). Technically, HCC 173 is not a CMSHCC because it is not included in the MA model. However, for expository purposes, we will refer to 71 CMSHCCs for the remainder of this report.
A relative weight is the incremental contribution of a particular health status marker to the risk score.

22

•

have no months of enrollment in a Medicare HMO during 2004;

•

have no months of working aged status in 2004;

•

be a U.S. resident during 2004; and

•

have at least one office or other outpatient evaluation and management (E&M) visit24
in 2004.

Expenditures are defined for risk adjustment model calibration as for the PGP
demonstration. The dependent variable for the regression model is annualized expenditures
capped at $100,000. All Medicare payments are incorporated into the dependent expenditure
variable. Regression models are weighted by the fraction of months during 2004 each beneficiary
is eligible for the sample. In addition to the sample exclusions listed above, the PGP concurrent
model was calibrated for beneficiaries without ESRD, and who were continuing enrollees.25
5.1.3 Relative Weights
The PGP concurrent model uses multiple regression analysis to estimate the incremental
expenditures associated with each CMS-HCC diagnostic category. When divided by national
average per capita Medicare expenditures, incremental expenditures may be expressed as a
“relative weight” for each CMS-HCC. For example, hypothetically, if the incremental
expenditures associated with HCC 80, Congestive Heart Failure, is $2,000 and national average
per capita Medicare expenditures are $5,000, then the relative weight for HCC 80 is 2,000/5,000
or 0.400. Relative weights represent the portion of a risk score associated with each of the model
variables. A risk score is created by summing the relative weights for markers assigned to a
beneficiary. Table 5 shows the relative weights for the variables included in the risk adjustment
model.

24

25

CPT codes used to identify Office or Other Outpatient E&M visits are as follows: 99201, 99202, 99203, 99204,
99205, 99211, 99212, 99213, 99214, and 99215.
Continuing enrollees in the calibration sample are enrolled in Medicare on January 1, 2004.

23

Table 5
PGP Concurrent Risk Adjustment Model for Continuing Enrollees Without ESRD
Variable
NOCMSHCC
HCC1
HCC2
HCC5
HCC7
HCC8
HCC9
HCC10
HCC15
HCC16
HCC17
HCC18
HCC19
HCC21
HCC25
HCC26
HCC27
HCC31
HCC32
HCC33
HCC37
HCC38
HCC44
HCC45
HCC51
HCC52
HCC54
HCC55
HCC67
HCC68
HCC69
HCC70
HCC71
HCC72
HCC73
HCC74
HCC75
HCC77
HCC78
HCC79
HCC80
HCC81
HCC82
HCC83
HCC92
HCC95

Label
No CMS-HCC2
HIV/AIDS
Septicemia/Shock
Opportunistic Infections
Metastatic Cancer and Acute Leukemia
Lung, Upper Digestive Tract, and Other Severe Cancers
Lymphatic, Head and Neck, Brain, and Other Major Cancers
Breast, Prostate, Colorectal and Other Cancers and Tumors
Diabetes with Renal or Peripheral Circulatory Manifestation
Diabetes with Neurologic or Other Specified Manifestation
Diabetes with Acute Complications
Diabetes with Ophthalmologic or Unspecified Manifestation2
Diabetes without Complication2
Protein-Calorie Malnutrition
End-Stage Liver Disease
Cirrhosis of Liver
Chronic Hepatitis
Intestinal Obstruction/Perforation
Pancreatic Disease
Inflammatory Bowel Disease
Bone/Joint/Muscle Infections/Necrosis
Rheumatoid Arthritis and Inflammatory Connective Tissue Disease
Severe Hematological Disorders
Disorders of Immunity
Drug/Alcohol Psychosis
Drug/Alcohol Dependence
Schizophrenia
Major Depressive, Bipolar, and Paranoid Disorders
Quadriplegia, Other Extensive Paralysis3
Paraplegia3
Spinal Cord Disorders/Injuries4
Muscular Dystrophy2
Polyneuropathy
Multiple Sclerosis
Parkinson's and Huntington's Diseases
Seizure Disorders and Convulsions
Coma, Brain Compression/Anoxic Damage5
Respirator Dependence/Tracheostomy Status
Respiratory Arrest
Cardio-Respiratory Failure and Shock
Congestive Heart Failure
Acute Myocardial Infarction
Unstable Angina and Other Acute Ischemic Heart Disease
Angina Pectoris/Old Myocardial Infarction
Specified Heart Arrhythmias
Cerebral Hemorrhage

24

Relative
Weight1
0.182
0.300
1.440
0.719
1.860
1.860
0.703
0.319
0.302
0.302
0.268
0.182
0.182
1.525
0.701
0.211
0.211
1.026
0.597
0.334
0.968
0.285
0.929
1.382
1.023
0.512
0.679
0.472
1.102
1.102
0.676
0.182
0.336
0.389
0.373
0.304
0.814
2.672
1.656
1.112
0.433
1.893
1.031
0.394
0.420
1.350
(continued)

Table 5 (continued)
PGP Concurrent Risk Adjustment Model for Continuing Enrollees Without ESRD
Variable
HCC96
HCC100
HCC101
HCC104
HCC105
HCC107
HCC108
HCC111
HCC112
HCC119
HCC130
HCC131
HCC132
HCC148
HCC149
HCC150
HCC154
HCC155
HCC157
HCC158
HCC161
HCC164
HCC173
HCC174
HCC176
HCC177

Label
Ischemic or Unspecified Stroke
Hemiplegia/Hemiparesis3
Cerebral Palsy and Other Paralytic Syndromes
Vascular Disease with Complications
Vascular Disease
Cystic Fibrosis
Chronic Obstructive Pulmonary Disease
Aspiration and Specified Bacterial Pneumonias
Pneumococcal Pneumonia, Emphysema, Lung Abscess
Proliferative Diabetic Retinopathy and Vitreous Hemorrhage2
Dialysis Status
Renal Failure
Nephritis2
Decubitus Ulcer of Skin
Chronic Ulcer of Skin, Except Decubitus2
Extensive Third-Degree Burns
Severe Head Injury5
Major Head Injury
Vertebral Fractures without Spinal Cord Injury4
Hip Fracture/Dislocation
Traumatic Amputation
Major Complications of Medical Care and Trauma
Major Organ Transplant (procedure)
Major Organ Transplant Status
Artificial Openings for Feeding or Elimination
Amputation Status, Lower Limb/Amputation Complications

Relative
Weight1
0.477
1.102
0.375
1.041
0.330
0.435
0.319
1.078
0.536
0.182
0.618
0.618
0.182
1.090
0.182
2.915
0.814
0.610
0.676
1.676
1.661
1.457
5.375
0.502
0.981
0.831

NOTES:
1
The incremental predicted expenditures from the regression model were converted to relative risk scores by
26
dividing by the sample national average of expenditures, $7,727.84. The relative weights from all HCCs
assigned to a beneficiary are summed to determine his/her risk score.
2
The relative weights of these HCCs and NOCMSHCC were constrained to be equal.
3
The relative weights of these HCCs were constrained to be equal.
4
The relative weights of these HCCs were constrained to be equal.
5
The relative weights of these HCCs were constrained to be equal.
SOURCE: RTI International analysis of 2004 Medicare 5% sample.

26

The entire national sample of beneficiaries eligible for the PGP demonstration is used to compute this average,
including new and continuing enrollees, and ESRD enrollees.

25

5.1.4 Constraints
Some of the regression coefficients for the PGP concurrent model were constrained to
ensure that incremental expenditure predictions and relative weights have certain desirable
properties (see Pope et al., 2004 for further discussion of model constraints). Clinical consultants
to CMS suggested that metastatic cancer is not consistently correctly recorded on Medicare
claims, so the relative weights for Metastatic Cancer and Acute Leukemia (HCC 7) and Lung,
Upper Digestive Tract, and Other Severe Cancers (HCC 8) were constrained to be equal.
In addition, the relative weights of several CMS-HCCs were constrained to equal the
relative weight of the NOCMSHCC variable because the unconstrained relative weights violate
the principle that providers should not be penalized for recording additional diagnoses. That is,
without constraint, a provider's risk score could be lower if it recorded one of the CMS-HCC
diagnoses. We therefore constrained 6 CMS-HCCs (HCC 18, HCC 19, HCC 70, HCC 119,
HCC132, and HCC 149) to have relative weights equal to the relative weight for the
NOCMSHCC variable.
Lastly, six sets of CMS-HCCs were constrained because the unconstrained relative
weights violate the principle that higher ranked conditions in a clinical disease hierarchy should
have higher predicted costs. Each of these three pairs were constrained to have equal relative
weights: HCC 15 and HCC16, HCC 26 and HCC 27, and HCC 130 and HCC131. Relative
weights for Quadriplegia, Other Extensive Paralysis, Paraplegia, and Hemiplegia/Hemiparesis
(HCCs 67, 68 and 100) were constrained to equal the relative weight for Quadriplegia, Other
Extensive Paralysis (HCC 67). Similarly, the relative weights for Coma, Brain
Compression/Anoxic Damage (HCC 75) was constrained to equal the relative weight for Severe
Head Injury (HCC 514). Lastly, Spinal Cord Disorders/Injuries (HCC 69) was constrained to
have an equal relative weight to Vertebral Fractures without Spinal Cord Injury (HCC 157).
5.2

Demographic Adjustment

A primary goal of risk adjustment for payment systems is to ensure that expenditures for
beneficiaries with observable characteristics are correctly predicted. To ensure that mean
predictions for beneficiaries by demographic subgroup are accurate, we created demographic
multipliers to adjust mean expenditure predictions for demographic categories to the actual
expenditure mean of each sub-population. The multipliers are calculated as the ratio of actual
mean expenditures for a subgroup to mean expenditures for a subgroup predicted from the
regression model described above.
Demographic modifiers were created for age, sex, and Medicaid status to ensure that on
average, these demographic groups are predicted correctly.27 Average predicted payments should
27

We also investigated an adjuster for "originally disabled" status, that is, beneficiaries currently entitled by age
who were originally entitled to Medicare by disability. This demographic factor is included in the prospective
CMS-HCC model used for MA plan payment. However, we found that after controlling for age, sex, and
Medicaid status, the incremental originally disabled adjuster appeared to be negligible and was difficult to
estimate precisely with available sample sizes. We did not include an adjuster for originally disabled status in the
final model.

26

equal average actual payments within each age/sex and Medicaid group. Modifiers adjust each
individual’s initial risk score multiplicatively based on their demographic information.
Beneficiary age was grouped into seven categories based on the age/sex cells used in the
prospective CMS-HCC model. Certain prospective model cells for older and younger
beneficiaries with relatively small sample sizes were merged to acquire stable modifiers. Each
demographic category we defined has sufficient sample size for creating an accurate modifier.
Table 6 shows the modifiers for each demographic category. There are seven age/sex categories
for males and females (0–54, 55–64, 65–69, 70–74, 75–79, 80–84, and 85+) in Medicaid and
non-Medicaid status, resulting in a total of 28 demographic modifiers (7 x 2 x 2 = 28). Each
beneficiary is assigned to one and only one demographic category.
Table 6
PGP Concurrent Risk Adjustment Model Demographic Modifiers
Multiplier
Medicaid
Non-Medicaid

Demographic Group
Female
0-54 Years
55-64 Years
65-69 Years
70-74 Years
75-79 Years
80-84 Years
85 Years or Over

1.012
1.025
1.061
1.063
1.048
1.043
1.025

0.946
0.965
1.001
1.010
1.007
0.987
0.980

Male
0-54 Years
55-64 Years
65-69 Years
70-74 Years
75-79 Years
80-84 Years
85 Years or Over

0.892
0.937
0.993
1.005
1.010
1.010
1.010

0.817
0.883
0.963
0.972
0.966
0.944
0.933

SOURCE: RTI International analysis of 2004 Medicare 5% sample.
5.3

Risk Score Calculation

CMS-HCC diagnostic categories contribute additively to expenditure prediction,
weighted by their expected incremental contribution to expenditures. A beneficiary assigned
multiple HCCs based on their claims history will receive the sum of the relative weights for
those HCCs as their initial risk score. The demographic adjuster is applied to the initial score to
produce the final risk score.
27

Consider the example from Section 2 of a 79 year-old female Medicaid enrollee who has
been diagnosed with AMI, angina, COPD, renal failure, and ankle sprain. Recall that the first
step of risk score calculation is to assign risk markers and a demographic category. This
beneficiary would be assigned CMS-HCCs for AMI (HCC 81), COPD (HCC 108), and renal
failure (HCC 131).28 This beneficiary would not receive the NOCMSHCC marker due to the
assignment of at least one CMS-HCC.
The next steps are to attach relative weights and calculate the initial risk score. Table 7
describes the relative weights and calculation of the initial risk score. Note that Angina Pectoris
(HCC 83) and Ankle Sprain (HCC 162) do not receive relative weights as they are not assigned
as risk markers for this beneficiary.
Table 7
Hypothetical Example of Initial Risk Score Calculation
AMI (HCC 81)
Angina pectoris (HCC 83)1
COPD (HCC 108)
Renal failure (HCC 131)
Ankle sprain (HCC 162)2
TOTAL

1.893
0.000
0.319
0.618
0.000
2.830
Initial Risk Score = 2.830

1

HCC 83, angina pectoris has an incremental prediction, but the amount is not added because HCC 81,
AMI, is within the same hierarchy and is the more severe manifestation of cardiovascular disease.

2

HCC 162, ankle sprain is excluded from the CMS-HCC list due to its low impact on expenditures.

SOURCE: RTI International

The initial risk score for this beneficiary is equal to 2.830. As a final step, the risk score is
modified by the appropriate demographic multiplier. The appropriate modifier for a 79 year-old
female Medicaid enrollee from Table 6 is 1.048 (Age 75-79, Female, Medicaid). The final risk
score is calculated as:
Final Risk Score = (Initial Risk Score) * (Demographic Modifier)

Final Risk Score = (2.830 * 1.048) = 2.966
This beneficiary’s final risk score would be 2.966. Compared to an average Medicare
enrollee eligible for the PGP demonstration, with a risk score of 1.000, this beneficiary is
expected to be almost three times as expensive.

28

See Section 2 for a full description of the CMS-HCC assignment process.

28

Note that beneficiaries with none of the 71 CMS-HCCs will be assigned an initial risk
score of 0.182, corresponding to the relative weight of the NOCMSHCC variable. Beneficiaries
with none of the significant diseases represented by the 71 CMS-HCCs are healthier than the
average Medicare beneficiary eligible for the demonstration, and are expected to use less than
20% of the health care services a Medicare beneficiary on average would be expected to use
during a year. Initial risk scores for beneficiaries without a CMS-HCC are also modified based
on the demographic category assigned to the beneficiary.
5.4

Summary

The PGP concurrent risk adjustment model uses diagnosis, procedure, and demographic
information to produce risk scores for aged/disabled continuing enrollees. Risk markers are
assigned and their relative weights summed to produce an initial risk score. That unmodified risk
score is then multiplied by the demographic multiplier to produce a final risk score. The final
risk score is used in the PGP demonstration to adjust expenditures for health risk.

29

[This page left intentionally blank.]

30

SECTION 6
NEW ENROLLEE AND ESRD MODELS
In the last section, we document the development of the PGP concurrent risk adjustment
model, which is applied to continuing, aged/disabled beneficiaries. In this section we describe
the model designed for newly enrolled Medicare beneficiaries, and the model designed for ESRD
beneficiaries. ESRD beneficiaries are subdivided into those currently treated with dialysis or a
kidney transplant, or those with functioning grafts. New Medicare enrollees are those not
enrolled in Medicare at the beginning of a performance year, and hence not having a full
diagnostic profile with which to produce risk markers. Both of these models are based on models
designed for and used by CMS for Medicare Advantage plan payments. The models have been
calibrated for the PGP demonstration sample in a similar fashion to the PGP concurrent model
presented in the preceding section. This section presents both models beginning with the PGP
new enrollee model.
6.1

PGP New Enrollee Model

New Medicare enrollees are defined as beneficiaries enrolled at least one month in both
Part A and Part B Medicare during a demonstration performance year, but not enrolled on the
first day of the performance year. These beneficiaries do not have a twelve month history of
diagnoses to generate a complete diagnostic profile. The PGP new enrollee model is therefore
based only on demographic information available at the time of enrollment. The model uses age,
gender, and Medicaid status to estimate expected expenditures.
6.1.1 Model Calibration and Variables
RTI calibrated the new enrollee model on the year 2004 5% national random sample of
beneficiaries, the same data used to calibrate the continuing enrollee model described in Section
5.29 The only difference in the sample was the inclusion of new enrollees in addition to
continuing enrollees. Ideally, the new enrollee model would have been calibrated on a sample of
new enrollees only. However, the vast majority of new Medicare enrollees are beneficiaries that
age into Medicare at 65 years of age. Because of this, our sample of new enrollees is heavily
weighted towards new enrollees who are 65 years of age. Although our calibration sample of age
65 new enrollees was sufficient to produce statistically reliable expenditure estimates for age 65
new enrollees, it was insufficient for the other age groups. To remedy this, the PGP new enrollee
model was calibrated on the merged sample including both continuing and new enrollee PGP
sample to gain enough sample size for the age groups above and below 65 years of age. The
implicit assumption is that expenditures for new and continuing enrollees are similar for most
ages, which prior analysis has shown to be a reasonable assumption (Pope et al., 2004).

29

New enrollees with dialysis months are not included in the sample and are not given risk scores from the new
enrollee model. See Section 6.2.2 for a discussion of how new enrollees undergoing dialysis treatment are given
risk scores.

31

The new enrollee model was calibrated by regressing total annualized expenditures
capped at $100,000 on a set of age/sex category variables and Medicaid status,30 for the
combined sample of new and continuing enrollees. The most common way to qualify for
Medicare is by age. Because of the large proportion of new enrollees who are 65 years of age,
separate relative weights are estimated for age 65 (and for ages 66, 67, 68, and 69). This allows
the age 65 relative weights to be more accurate. The age-sex cells in the PGP new enrollee model
are the same as those used in the CMS-HCC demographic model for new enrollees (Pope et al.,
2004). In addition to age-sex cells, the PGP new enrollee model includes relative weights based
on Medicaid status. The incremental Medicaid relative weights are differentiated by 10 age-sex
ranges, including a separately estimated Medicaid effect for age 65.
A risk score for each mutually exclusive demographic category (age/sex and Medicaid)
was derived from the regression model estimated coefficients and is shown in Table 8. The risk
score expresses predicted expenditures relative to the national mean expenditure in 2004.
To illustrate, consider a male beneficiary (New Enrollee A) that enrolls in Medicare at
age 65. The beneficiary will receive a risk score of 0.646, compared to a Medicaid dual-eligible
beneficiary of the same age and sex (New Enrollee B) who would receive a risk score of 1.235.
Each beneficiary is assigned one risk score based on their age, sex, and Medicaid status.
6.1.2 Adjustment to Predict New Enrollee Mean Expenditures Accurately
The new enrollee regression model predicts the overall mean expenditures accurately for
the merged sample of continuing and new enrollees used to estimate this model. It does not
predict the correct mean for the new enrollee sub-population alone. Specifically, it underpredicts
expenditures for the new enrollee sub-population by 1.1%. To predict new enrollee mean
expenditures correctly, all beneficiaries receiving a risk score from the PGP new enrollees model
are subject to a “multiplier” of 1.011 that scales expenditure predictions to the actual new
enrollees mean. Continuing the example begun in Section 6.1.1, Figure 3 provides an illustration
of the application of the overall multiplier for the PGP new enrollee model.
6.1.3 Summary
The PGP new enrollee model provides an accurate prediction for beneficiaries that are
new to Medicare without relying on an incomplete diagnosis profile. Beneficiary demographic
characteristics available at the time of enrollment are all that is required to generate risk scores.
This model is applied to all aged/disabled beneficiaries that are not enrolled in Medicare at the
beginning of a base or performance year.
6.2

PGP ESRD Model

Beneficiaries with end stage renal disease (ESRD) are treated with dialysis and kidney
transplants. To more precisely account for the higher average expenditures of Medicare

30

We did not include originally disabled status among the predictive factors for the PGP new enrollee model
because new Medicare enrollees are rarely in originally disabled status (by definition, a beneficiary cannot be
originally disabled when he/she first enrolls in the Medicare program).

32

Table 8
PGP Demographic Model for New Enrollees1 Initial Risk Scores
Risk Score2
Non-Medicaid
Medicaid3
Female
0-34 Years
35-44 Years
45-54 Years
55-59 Years
60-64 Years
65 Years
66 Years
67 Years
68 Years
69 Years
70-74 Years
75-79 Years
80-84 Years
85-89 Years
90-94 Years
95 Years or Over
Male
0-34 Years
35-44 Years
45-54 Years
55-59 Years
60-64 Years
65 Years
66 Years4
67 Years4
68 Years
69 Years
70-74 Years
75-79 Years
80-84 Years
85-89 Years
90-94 Years
95 Years or Over

0.587
0.697
0.843
0.943
1.029
0.556
0.582
0.611
0.628
0.651
0.731
0.877
0.991
1.110
1.210
1.264

0.857
0.967
1.113
1.213
1.299
1.137
1.139
1.168
1.185
1.208
1.250
1.348
1.462
1.581
1.681
1.735

0.442
0.646
0.785
0.930
1.064
0.646
0.687
0.687
0.745
0.767
0.870
1.048
1.194
1.332
1.412
1.510

0.725
0.929
1.068
1.213
1.347
1.235
1.276
1.276
1.334
1.356
1.459
1.578
1.724
1.862
1.942
2.040

NOTES:
1
Aged and disabled beneficiaries. Excludes ESRD and working aged beneficiaries.
2
The predicted dollar amounts from the regression were converted to risk scores by dividing by the
sample national average of expenditures, $7,727.84. Note that each category is mutually exclusive
and therefore the relative weight for each category is presented as a risk score.
3
Medicaid male beneficiaries 65 years of age were constrained to have their Medicaid coefficient
equal to the Medicaid coefficient for male beneficiaries 66–69 years of age and male beneficiaries 70
to 74 years of age.
4
Male beneficiaries aged 66 were constrained to have their coefficients equal to male beneficiaries 67
years of age.

33

Figure 3
Application of the Overall Multiplier for the PGP New Enrollees Model
New Enrollee Multiplier1

1.011

Therefore the final risk scores for New Enrollees A and B would be:
New Enrollee A: Initial Risk Score = 0.646
Final Risk Score = 0.646 * 1.011 = 0.653
New Enrollee B: Initial Risk Score = 1.235
Final Risk Score = 1.235 * 1.011 = 1.249
NOTES:
1

Mean predicted expenditures for the new enrollees sub-population equals $5,541, and actual mean
expenditures equals $5,603. The overall multiplier thus equals $5,603 / $5,541 = 1.011.

SOURCE: RTI International analysis of 2004 Medicare 5% sample.

beneficiaries with ESRD, RTI designed a separate concurrent risk adjustment model for ESRD
beneficiaries. The PGP ESRD model is based on the prospective CMS-HCC ESRD model
developed by CMS and currently used to set payment rates for ESRD beneficiaries enrolled in
Medicare Advantage plans.31 The PGP ESRD model is comprised of three submodels, including
different treatments for dialysis, transplant, and functioning graft beneficiaries. The model
adjusts based on the individual’s actual course of treatment during a base or performance year.
The PGP ESRD model has been calibrated for the PGP demonstration sample in a similar
fashion to the PGP concurrent and new enrollee models presented in preceding sections.
6.2.1 Defining ESRD Beneficiaries
ESRD beneficiaries are identified by their enrollment and ESRD information recorded by
Medicare. Beneficiaries identified by Medicare as having dialysis treatments or a kidney
transplant32 during a payment or base year are counted as ESRD beneficiaries. In addition,
Medicare beneficiaries with kidney transplants in prior years are identified as ESRD
beneficiaries. Medicare maintains an enrollment database (EDB) documenting dialysis and
kidney transplant dates for all ESRD beneficiaries. This data is used to identify all ESRD
beneficiaries.
The PGP ESRD model risk score depends on the number of months spent in each ESRD
status during an analysis year. ESRD beneficiaries are treated with dialysis and kidney
transplants, and the course of treatment determines the costs incurred by the beneficiary during
31

32

The CMS-HCC ESRD prospective risk adjustment model is described in the following CMS 45-day notice:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/Downloads/Advance2005.pdf. The final model
coefficients are presented in this 2005 announcement:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/downloads/Announcement2005.pdf.
Including a simultaneous Kidney/Pancreas transplant.

34

the year. Beneficiaries are assigned to one of three ESRD statuses for every month of a base or
performance year. An ESRD beneficiary can be categorized as a transplant, dialysis, or
functioning graft beneficiary in any given month (certain ESRD beneficiaries will also have
some months not spent in any ESRD status).
Transplant beneficiaries are identified by the date of their transplant surgery, and are
included as transplant beneficiaries for the month of surgery and the two months following that
surgery. Dialysis beneficiaries are identified by the dialysis start and end dates on the enrollment
file. Functioning graft beneficiaries are enrollees that have had a kidney transplant, and are not
currently being treated with dialysis.
An ESRD beneficiary’s final risk score is dependent on:
•

the number of months spent outside of ESRD status (i.e., aged/disabled);

•

the number of months a beneficiary is treated with dialysis;

•

whether a beneficiary receives a kidney transplant; and

•

the number of months a beneficiary is “functioning graft,” and how many months the
beneficiary is post-transplant.

Dialysis Beneficiaries
Dialysis start and end dates recorded on the EDB define each beneficiary’s dialysis
status. A beneficiary with a dialysis start date is assigned to dialysis status beginning the first of
the month after that dialysis start date. A beneficiary is continuously assigned to dialysis status
until a dialysis end date is recorded on the EDB, or the beneficiary dies.33 A beneficiary with
both a transplant and dialysis period covering the same month will be assigned to transplant
status for that month. A beneficiary with a dialysis start date of May 15, and dialysis end date of
July 15 but no transplant will be assigned to dialysis status for June and July. A beneficiary with
a dialysis start date of May 15 and transplant start date of July 15 will be assigned to dialysis
status for June and transplant status for July.
Transplant Beneficiaries
Medicare records the date of each kidney transplant for beneficiaries with ESRD. Kidney
transplants are associated with much higher medical expenditures for the month of transplant and
the two following months. Therefore, all beneficiaries with a transplant record on the EDB are
assigned to transplant status for the month of transplant and the two following months, or until
the death of the beneficiary. A beneficiary with a transplant date of, for example, May 15 will be
assigned to transplant status for May, June, and July.

33

Beneficiaries that die during the demonstration do not have their months included in any calculations beginning
the first of the month after the date of death.

35

Functioning Graft Beneficiaries
Beneficiaries who have undergone kidney transplant surgery and do not require dialysis
or another transplant are considered “Functioning Graft” beneficiaries. These beneficiaries are
assumed to have a working kidney transplant. Post-transplant beneficiaries are less expensive
than either dialysis or transplant beneficiaries. Expenditures for functioning graft patients remain
high, but have a cost pattern that is closer to the general population than to dialysis patients.
Beneficiaries identified as functioning graft maintain that status from the fourth month posttransplant until they return to dialysis status, receive another kidney transplant, or die. A
beneficiary with a transplant date of, for example, May 15 will be assigned to functioning graft
status from August onwards.
Functioning graft beneficiaries are further delineated into two categories to better address
the decreasing cost pattern after the transplant. The first months after transplant are associated
with higher costs, but as the patient recovers, the higher costs of these beneficiaries are driven
primarily by the Part B covered immunosuppressive drugs. The first functioning graft category,
category I, covers the fourth month post-transplant through the tenth. This category covers the
higher service intensity during this period. The second functioning graft category, category II,
covers the eleventh month post-transplant onward. Functioning graft status is assigned as of the
most recent transplant, as beneficiaries may undergo more than one transplant. A beneficiary that
has transplant dates of May 15 and June 15 will be assigned to functioning graft category I in
September and functioning graft category II in April of the following year.
6.2.2 PGP ESRD Dialysis Model
Dialysis beneficiaries incur monthly costs for dialysis and are more expensive across the
entire spectrum of disease than beneficiaries entitled to Medicare by age or disability. Dialysis
beneficiaries average close to $60,000 in annual expenditures compared to aged/disabled
beneficiaries who have mean costs closer to $7,000 annually. To account for these higher
expenditures a separate concurrent risk adjustment model was developed for beneficiaries
identified in dialysis status.
The PGP ESRD model for dialysis patients is similar to the PGP concurrent risk
adjustment model for aged/disabled enrollees. The PGP ESRD model for dialysis patients also
uses CMS-HCCs to estimate health expenditures, but changes were made to incorporate
differences between the ESRD and aged/disabled populations.
Certain CMS-HCCs were not included in the PGP ESRD model (see Table 1 for the list
of CMS-HCCs). Dialysis Status (HCC 130), Renal Failure (HCC 131), and Nephritis (HCC 132)
are excluded because they are conditions that have a lower ranking in the disease hierarchy than
ESRD Dialysis Status (HCC 129), which all beneficiaries in the PGP ESRD dialysis model must
have. The remaining 68 CMS-HCCs are included in the model.
Age-sex terms are included in the PGP ESRD dialysis model as interactions with dialysis
status. There is no second stage adjustment for age, sex, and Medicaid status as occurs in the
aged-disabled model. A total of eight age-sex interactions with dialysis status are included in the

36

model, two sets of four age groups for male and female (less than 55, 55 to 64, 65 to 74 years of
age, and greater than or equal to 75 years of age).
The PGP ESRD dialysis model was created using a combined sample of aged/disabled
and dialysis enrollees. Although estimating the dialysis model on a sample of ESRD
beneficiaries alone would have been preferred, we did not have enough ESRD beneficiaries in
our 5% national random sample to do so. The combined sample ensures that each HCC has
sufficient sample size to generate an accurate prediction. The sample exclusions applied for this
model were the same as those applied for the sample used to create the PGP concurrent model
described in a previous section (see Section 5.1.2).
The combined sample model predicts mean expenditures for ESRD beneficiaries
accurately (because of the inclusion of the age/sex intercepts), and allows some adjustment for
their diagnostic profile. The estimated regression coefficients for the HCCs diagnostic categories
in the model are very similar to the coefficients estimated for the aged/disabled model presented
in Section 5 because aged/disabled beneficiaries account for 99 percent of the combined sample.
Ideally the HCC coefficients would be customized for the ESRD population, but this was not
feasible because of the small available sample size of ESRD beneficiaries in our data.
The relative weights for the PGP ESRD dialysis model are shown in Table 9.
The PGP ESRD dialysis model is an additive model like the PGP concurrent risk
adjustment model for aged/disabled beneficiaries. A dialysis beneficiary that was in dialysis
status for an entire year would receive a risk score equal to the sum of the relative weight for the
beneficiary's age-sex cell and the relative weights for the HCCs the beneficiary was diagnosed
with during the year. Alternatively, a beneficiary with both dialysis months and aged/disabled
months would have a final risk score equal to the weighted average of their aged/disabled and
dialysis risk scores (weighted by the number of months spent in aged/disabled versus dialysis
status). For example, if a beneficiary had 3 months of aged/disabled eligibility with an
aged/disabled risk score of 2.000, and 9 months of dialysis treatment with a dialysis risk score of
10.000, then the beneficiary's final risk score would be 8.00034.
New enrollees that are identified as having been treated with dialysis will not be given
risk scores from the model above. New enrollees do not have a complete diagnostic profile to
generate a risk score. These beneficiaries will be given an initial risk score equal to the average
annualized payment for dialysis beneficiaries, 7.61735. This initial risk score is not modified for
any demographic characteristics, but will be weighted by the number of months the enrollee is
assigned to dialysis status.

34
35

(0.25*2.000) + (0.75*10.000) = 8.000
The mean annualized expenditure for dialysis beneficiaries from the PGP 2004 sample is equal to $58,865.35.
This was converted to a risk score by dividing by the national average expenditures for all beneficiaries,
$7,727.84.

37

Table 9
PGP ESRD Dialysis Model
Relative
Weight1

Variable
Female
Age Less Than 55
Age 55 to 64
Age 65 to 74
Age 75 or Greater
Male
Age Less Than 55
Age 55 to 64
Age 65 to 74
Age 75 or Greater
Diseases
HCC1
HCC2
HCC5
HCC7
HCC8
HCC9
HCC10
HCC15
HCC16
HCC17
HCC18
HCC19
HCC21
HCC25
HCC26
HCC27
HCC31
HCC32
HCC33
HCC37
HCC38

4.004
3.904
3.995
4.064
3.974
3.624
3.813
3.789
HIV/AIDS
Septicemia/Shock
Opportunistic Infections
Metastatic Cancer and Acute Leukemia
Lung, Upper Digestive Tract, and Other Severe Cancers
Lymphatic, Head and Neck, Brain, and Other Major Cancers
Breast, Prostate, Colorectal and Other Cancers and Tumors
Diabetes with Renal or Peripheral Circulatory Manifestation
Diabetes with Neurologic or Other Specified Manifestation
Diabetes with Acute Complications

0.325
1.424
0.717
1.861
1.861
0.707
0.318
0.317
0.317
0.262

Diabetes with Ophthalmologic or Unspecified Manifestation2

0.181

2

Diabetes without Complication
Protein-Calorie Malnutrition
End-Stage Liver Disease
Cirrhosis of Liver
Chronic Hepatitis
Intestinal Obstruction/Perforation
Pancreatic Disease
Inflammatory Bowel Disease
Bone/Joint/Muscle Infections/Necrosis
Rheumatoid Arthritis and Inflammatory Connective Tissue Disease

38

0.181
1.504
0.717
0.229
0.229
1.010
0.606
0.334
0.948
0.285
(continued)

Table 9 (continued)
PGP ESRD Dialysis Model
Relative
Variable
HCC44
HCC45
HCC51
HCC52
HCC54
HCC55
HCC67
HCC68
HCC69
HCC70

Severe Hematological Disorders
Disorders of Immunity
Drug/Alcohol Psychosis
Drug/Alcohol Dependence
Schizophrenia
Major Depressive, Bipolar, and Paranoid Disorders
Quadriplegia, Other Extensive Paralysis3
Paraplegia3
Spinal Cord Disorders/Injuries4

HCC71
HCC72
HCC73
HCC74
HCC75
HCC77
HCC78
HCC79
HCC80
HCC81
HCC82
HCC83
HCC92
HCC95
HCC96
HCC100
HCC101
HCC104
HCC105
HCC107
HCC108
HCC111
HCC112

Muscular Dystrophy2
Polyneuropathy
Multiple Sclerosis
Parkinson's and Huntington's Diseases
Seizure Disorders and Convulsions
Coma, Brain Compression/Anoxic Damage5
Respirator Dependence/Tracheostomy Status
Respiratory Arrest
Cardio-Respiratory Failure and Shock
Congestive Heart Failure
Acute Myocardial Infarction
Unstable Angina and Other Acute Ischemic Heart Disease
Angina Pectoris/Old Myocardial Infarction
Specified Heart Arrhythmias
Cerebral Hemorrhage
Ischemic or Unspecified Stroke
Hemiplegia/Hemiparesis3
Cerebral Palsy and Other Paralytic Syndromes
Vascular Disease with Complications
Vascular Disease
Cystic Fibrosis
Chronic Obstructive Pulmonary Disease
Aspiration and Specified Bacterial Pneumonias
Pneumococcal Pneumonia, Emphysema, Lung Abscess

HCC119

Proliferative Diabetic Retinopathy and Vitreous Hemorrhage2

39

Weight1
0.919
1.364
1.014
0.517
0.681
0.474
1.097
1.097
0.676
0.202
0.337
0.387
0.372
0.305
0.768
2.585
1.593
1.120
0.443
1.885
1.030
0.391
0.423
1.347
0.476
1.097
0.374
1.048
0.335
0.454
0.318
1.063
0.539
0.181
(continued)

Table 9 (continued)
PGP ESRD Dialysis Model
Relative
Variable
HCC148
HCC149
HCC150
HCC154
HCC155
HCC157
HCC158
HCC161
HCC164
HCC173
HCC174
HCC176
HCC177

Weight1
1.078

Decubitus Ulcer of Skin
Chronic Ulcer of Skin, Except Decubitus2
Extensive Third-Degree Burns
Severe Head Injury5
Major Head Injury
Vertebral Fractures without Spinal Cord Injury4
Hip Fracture/Dislocation
Traumatic Amputation
Major Complications of Medical Care and Trauma
Major Organ Transplant (procedure)
Major Organ Transplant Status
Artificial Openings for Feeding or Elimination
Amputation Status, Lower Limb/Amputation Complications

0.181
2.942
0.768
0.609
0.676
1.661
1.543
1.460
4.808
0.448
0.971
0.859

NOTES:
1
The dollar amounts in this table were converted to relative risk scores by dividing by the national
average of expenditures, $7,727.84.
2
These HCCs were constrained to equal the coefficient for NOCMSHCC. Note that ESRD
beneficiaries can not receive the NOCMSHCC variable, as ESRD is considered a
significant condition.
3
These HCCs were constrained to have equal coefficients.
4
These HCCs were constrained to have equal coefficients.
5
These HCCs were constrained to have equal coefficients.
SOURCE: RTI International analysis of 2004 Medicare 5% sample.

6.2.3 Transplant Adjustment
Beneficiaries that undergo a kidney transplant operation are treated differently from
dialysis ESRD beneficiaries when calculating a risk score. A kidney transplant incurs a high
dollar amount that does not vary drastically from patient to patient in a systematic way. The cost
pattern for a transplant beneficiary reflects the high inpatient costs associated with the transplant
surgery itself, as well as the higher service intensity for the 2 months after a transplant occurs.
Relative weight adjustments for the month of transplant and the two months following were
created from the average costs of these beneficiaries as estimated by CMS researchers.

40

An ESRD beneficiary that has a kidney transplant has the first month relative weight
weighted into their risk score.36 The same holds true for months 2 and 3, though the relative
weight is lower for the second and third months. Table 10 shows the transplant relative weight
adjustments for each month of transplant. A beneficiary surviving the three months of transplant
would receive an addition of 86.726 weighted into their final risk score, reflecting the
extraordinarily high costs of kidney transplant operations and follow-up treatment.37 These
transplant relative weights are weighted into the final risk score based on their total months
eligibility as described in Section 6.2.5.
Table 10
PGP ESRD Model—Transplant Relative Weights
Kidney Transplant
Month 1 Relative Weight1
Month 2 Relative Weight2
Month 3 Relative Weight3

68.256
9.235
9.235

NOTES:
1
Transplant payments are taken from the CMS MA payment ESRD model. See CMS website:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/downloads/Announcement2005.pdf, accessed
January 2005. We derived a PGP demonstration relative weight as follows, accounting for the relative
frequency and cost of kidney (95.1%) versus kidney/pancreas transplants (4.9%), mean 2004 dialysis
expenditures of $68,556.27 (the CMS transplant factors are relative to year 2000 mean dialysis
expenditures) and our PGP sample average costs of $7,727.84. Transplant Month 1 = {[(7.510* 0.951)
+ (11.266 * 0.049) * 68,556.27]/7,727.84} = (527,474.96/7,727.84) = 68.256
2
Transplant payments are taken from the CMS MA payment ESRD model. See CMS website:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/downloads/Announcement2005.pdf, accessed
January 2005. We derived a PGP demonstration relative weight as follows, accounting for the relative
frequency and cost of kidney (95.1%) versus kidney/pancreas transplants (4.9%), mean 2004 dialysis
expenditures of $68,556.27 (the CMS transplant factors are relative to year 2000 mean dialysis
expenditures) and our PGP sample average costs of $7,727.84.
Transplant Month 2 = {[(1.016 * 0.951) + (1.525 * 0.049) * 68,556.27]/ 7,727.84} =
(71,363.03/7,727.84) = 9.235
3
Transplant payments are taken from the CMS MA payment ESRD model. See CMS website:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/downloads/Announcement2005.pdf, accessed
January 2005. We derived a PGP demonstration relative weight as follows, accounting for the relative
frequency and cost of kidney (95.1%) versus kidney/pancreas transplants (4.9%), mean 2004 dialysis
expenditures of $68,556.27 (the CMS transplant factors are relative to year 2000 mean dialysis
expenditures) and our PGP sample average costs of $7,727.84.
Transplant Month 2 = {[(1.016 * 0.951) + (1.525 * 0.049) * 68,556.27]/ 7,727.84} =
(71,363.03/7,727.84) = 9.235

36

37

See Section 6.2.5 for a description of the weighting process used to create the final risk score for beneficiaries
with ESRD.
Note that this risk score would be weighted into the final risk score according to the process described in Section
6.2.5. If the beneficiary were eligible for Medicare for the full 12 months, the 68.256 would receive a 1/12
weight, and the 9.235 would receive a 2/12 weight. Please see Section 6.2.5 for a more thorough review of the
final risk score calculation.

41

To illustrate the transplant adjustment, consider a beneficiary on dialysis with a risk score
of 10.000 who also has a complete transplant period. Assume the beneficiary spent 9 months in
dialysis, and received a transplant on October 1. The initial risk score from the PGP ESRD
dialysis model (10.000) is weighted by the fraction of the year spent in dialysis status (9/12 or
0.75). The transplant adjustments, 68.256 and 9.235, are weighted by the fraction of the year
spent in each transplant status (1/12 and 2/12 respectively) and then the transplant relative weight
adjustments are weighted in to the initial risk score. The final risk score for this beneficiary is
14.72738
6.2.4 Functioning Graft Adjustment
Beneficiaries who have undergone kidney transplant surgery and do not require dialysis
or another transplant are considered ‘Functioning Graft’ beneficiaries. These beneficiaries are
assumed to have a working kidney transplant and are therefore less expensive than beneficiaries
in other ESRD statuses. CMS estimated payments for these beneficiaries for the fourth through
thirty-sixth month after the transplant was performed and found that functioning graft patients
are more similar to the general aged/disabled population than to dialysis patients.39 The
functioning graft adjustment is therefore an adjustment to the PGP concurrent risk adjustment
model, rather than the PGP ESRD dialysis model.
Functioning graft patients have a recognizable cost pattern based on the number of
months the beneficiary is post-transplant. Costs immediately after transplant are relatively high
but decline rapidly to a stable average by month 11. For this reason, two sets of relative weight
adjustments were developed. The first relative weight adjustment is for the fourth through tenth
month after the transplant was performed. Recall that the first three months (including the month
of transplant) are treated as transplant months. Beneficiaries assigned to functioning graft status
for the fourth through tenth month after transplant receive a substantial add-on to their
aged/disabled risk score based on their age. The add-ons are smaller thereafter.
An adjustment is given to these beneficiaries to cover the additional costs of Part B
immunosuppressive drugs covered by Medicare and additional services they receive to monitor
and maintain the graft. Table 11 describes the relative weight adjustments for functioning graft
beneficiaries.
The functioning graft relative weight adjustment is an addition to the aged/disabled risk
score (PGP concurrent risk adjustment model–Table 5) that reflects the cost of Part B
immunosuppressive drugs. To illustrate, consider a 65-year-old beneficiary identified as
Functioning Graft II (i.e., post-transplant months 11 or more) for an entire year. The beneficiary
would receive an increase in their risk score of 1.691 (Table 11). Therefore a beneficiary with a
risk score of 1.000 from the PGP concurrent risk adjustment model would receive a final risk
score of 2.691 under the above assumptions.
38
39

(0.75 * 10.000) + (68.256 * (1/12)) + (9.235 * 2/12) = 14.727.
According to CMS ESRD research.
See CMS website: http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/Downloads/Advance2005.pdf,
accessed March 2005.

42

Table 11
PGP ESRD Model—Functioning Graft Adjustment1
Functioning Graft I - Post-Transplant Months 4 to 10
Beneficiaries < 65

3.091

Beneficiaries 65+

3.425

Functioning Graft II - Post-Transplant Months 11+
Beneficiaries < 65
Beneficiaries 65+

1.620
1.691

NOTES:
1

Functioning graft factors are taken from the CMS MA payment ESRD model. See CMS website:
http://www.cms.hhs.gov/MedicareAdvtgSpecRateStats/downloads/Announcement2005.pdf, accessed January
2005.

6.2.5 PGP ESRD Model Risk Score Calculation
Calculating a risk score for an ESRD beneficiary depends on the number of months a
beneficiary spends in each status. As an example, consider a male beneficiary, 72 years of age,
that begins the year enrolled in Medicare, qualifying through age. The beneficiary spends three
months in this status, before being diagnosed with ESRD, and undergoing dialysis treatment.
From April through July, the beneficiary is treated with dialysis and then undergoes a kidney
transplant in August. After recovering from the transplant the beneficiary is treated as a
functioning graft beneficiary for the remainder of the year based on the record indicating no
additional transplant or dialysis treatment. Over the year, the beneficiary is diagnosed with Renal
Failure (HCC 131), Vascular Disease with Complications (HCC 104), and Diabetes with Renal
Manifestation (HCC 15). This beneficiary’s assignment is shown in Figure 4.
Figure 4
Hypothetical ESRD Status Assignment
Beneficiary
begins dialysis
treatment

Beneficiary
has
transplant

Jan

Dec

Aged-Disabled
Status
Dialysis Status
Transplant Status
Functioning Graft I Status

SOURCE: RTI International

43

The assignment of months for this hypothetical beneficiary is shown in Table 12.
Table 12
Hypothetical Example of ESRD Monthly Assignment
Aged-Disabled Months

3

Dialysis Months

4

Transplant Months

3

Functioning Graft Months I

2

To calculate the final risk score it is first necessary to calculate risk scores from the PGP
concurrent risk adjustment model and the PGP ESRD Model. Recall from Table 5 that the initial
risk score for this example beneficiary equals 1.96140. The beneficiary receives a demographic
modifier of 0.972 (age 70–74, male, non-Medicaid)41, resulting in an aged/disabled risk score of
1.90642. For the months the beneficiary has been identified as functioning graft I, the relative
weight adjustment produces a risk score of 5.331 (1.906 + 3.425)43. Further, the PGP ESRD
dialysis model (Table 9) produces a risk score of 5.178 for this beneficiary44. Lastly, the first
month transplant risk score is 68.256, and the following two months are 9.235 (Table 10).
The next step to calculate the overall risk score is to take the weighted average of the
individual risk scores. The weight for each score is equal to the number of months out of 12 to
get an annual figure.
Risk Score = (1.906 * 3/12) + (5.178 * 4/12) + (68.256 * 1/12) + (9.235 * 2/12) + (5.331
* 2/12)

Final Risk Score = 10.318
6.2.6 Summary
The PGP ESRD Model is comprised of a separate dialysis model and adjustments for
functioning graft and transplant beneficiaries. This model depends on the assignment of
beneficiaries into the three ESRD statuses by month. The final risk score for an ESRD
beneficiary depends on the months of aged/disabled eligibility, as well as the months spent in
each of the ESRD statuses of dialysis, transplant, and functioning graft.

40
41
42
43
44

Equal to the sum of HCCs assigned: 0.302 (HCC15) + 1.041 (HCC104) + 0.618 (HCC131) = 1.961.
See Table 6.
1.961 * 0.972 = 1.906.
Functioning graft factor for Graft Type I beneficiaries, Aged ≥ 65 from Table 11.
Equal to the sum of markers assigned: 3.813 (Male, Age 70-74) + 0.317 (HCC15) + 1.048 (HCC104) + 0.000
(HCC131) = 5.178.

44

SECTION 7
DATA REQUIREMENTS & MODEL UPDATES
7.1

Data Requirements

For the PGP demonstration, diagnosis data will be taken from claims (bills) submitted by
Medicare fee-for-service providers for reimbursement. These will include claims from the
participating PGPs and from nonparticipating providers providing services to beneficiaries
assigned to participating PGPs. Participating providers are not required to submit any additional
data for risk adjustment beyond their normal fee-for-service claims to Medicare. ICD-9-CM
diagnosis codes and demographics are the primary inputs of the CMS-HCC risk adjustment
models. CPT procedure codes used to identify transplant patients and a few other high-cost
patient types are taken from physician bills only (hospital bills will not be used to identify
procedures). Diagnosis codes will be taken from the following four claim sources:
•

inpatient hospital claims;

•

hospital outpatient claims;

•

physician claims; and

•

clinically-trained non-physician claims.

Diagnoses submitted by sources not in this list (e.g., home health agencies) may be of
questionable accuracy.
Diagnostic coding completeness and accuracy is important for accurate risk adjustment.
For example, if a PGP manages an assigned beneficiary so as to avoid an unnecessary
hospitalization, the same ICD-9-CM diagnostic markers need to be recorded by one of the
accepted sources so that the health status of the beneficiary is accurately measured. Specifically,
suppose that an admission for an assigned beneficiary with congestive heart failure is avoided.
Congestive heart failure needs to be recorded as a diagnosis on a hospital outpatient or physician
claim sometime during the performance year so that the actual health status risk of this
beneficiary is measured.
It is important to note that chronic diagnoses need to be recorded at least once for each
beneficiary in every performance year. The system has no “memory.” But recording the same
diagnosis more than once in the same year has no effect on risk adjustment. Also, recording
diagnoses not included in the CMS-HCC model does not affect risk adjustment. Diagnoses are
not differentiated by setting--no greater health risk is assigned for an inpatient diagnosis than one
from a physician's office. Also, the time of year a diagnosis is recorded does not matter.
Finally, Medicare enrollment information available to CMS is used to assign age, sex,
and Medicaid status markers. They are also used to calculate risk scores. Those data are available
from the Medicare enrollment files.

45

7.2

Model Updates

Throughout the demonstration, RTI will add newly implemented ICD-9-CM diagnosis
codes and CPT procedure codes to the PGP risk adjustment model for the purposes of identifying
CMS-HCCs for beneficiaries. RTI will use new codes identified by CMS annually for updates.
The PGP risk adjustment model will not be recalibrated during the course of the threeyear demonstration.
7.3

Upward Trend in Risk Scores

It is likely that the average risk scores of beneficiaries assigned to the physician groups
participating in the PGP demonstration will rise over time, independent of any actual increase in
health status risk. An upward trend in national fee-for-service risk scores has been observed over
time, presumably due to more complete coding of diagnoses on claims. But average risk scores
of PGP comparison groups are expected to rise at the same rate. If risk scores for PGP assigned
beneficiaries and comparison groups rise at the same rate over time due to more complete
diagnostic coding, PGP performance payments will be unaffected. Thus, no adjustment for the
nationwide upward trend in risk scores over time will be made in the PGP demonstration.

46

SECTION 8
CONCLUSION
The concurrent CMS-HCC risk adjustment model accounts for approximately 50% of the
variation in health care expenditures among Medicare beneficiaries. Using concurrent risk
adjustment in the PGP Demonstration provides an accurate assessment of changes in the health
status of beneficiaries.

47

[This page left intentionally blank.]

48

REFERENCES
Pope, GC; Ellis, RP; Ash, AS; Ayanian, JZ; Bates, DW; Burstin, H; Iezzoni, LI; Marcantonio, E;
Wu, B: “Diagnostic Cost Group Hierarchical Condition Category Models for Medicare Risk
Adjustment” Final Report to the Health Care Financing Administration under Contract
No. 500-95-048, December 21, 2000. Health Economics Research, Inc., Waltham, MA.
Pope, GC; Kautter, J; Ingber, MJ; Levy, JM; Robst, J; Ellis, RP; Ash, AS: “Risk Adjustment of
Medicare Capitation Payments Using the CMS-HCC Model” Health Care Financing Review
25(4):119-141. Summer 2004.
Kautter, J; Pope, GC; Trisolini, M; et al.: “Physician Group Practice Bonus Methodology
Specifications” Report to the Centers for Medicare and Medicaid Services under Contract No.
500-00-0024, T.O. No. 13, December, 2004. RTI International, Waltham, MA.

49