Topic 3

Question 1

Questions to answer when building a clinical identification algorithm

Answer 1

A clinical algorithm is a set of rules that is applied to claims data set to identify the conditions present in the population

1. Where are the (D)iagnoses recorded?
2. What is the (S)ource of the diagnosis (claims, medical charts, etc)
3. If the source is claims, what claims should be (C)onsidered (inpatient, outpatient, laboratory, etc)
4. If the claim contains more than one diagnosis, (H)ow many diagnoses will be considered for identification
5. Over what (T)ime span, and how often, will a diagnosis have to appear in claims for that diagnosis to be incorporated
6. What procedures may be useful for determining (S)everity of a diagnosis?
7. What prescription (D)rugs may be used to identify conditions?

Mnemonic: DCD H T S S (DeCiDe How To Set Standards for the Algorithm)

Duncan Chapter 4 (Risk), Page 91

Question 2

Challenges when constructing a condition-based model

Answer 2

N - Large (N)umber of procedure and drug codes
S - Deciding the (S)everity level at which to recognize the condition
C - The impact of (C)o-morbidities for conditions that are often found together
U - The degree of (U)ncertainty with which the diagnosis has been identified
E - The (E)xtent of the data (claims data will cover all members, but self-reported data will not)
B - The type of (B)enefit design that underlies the data

Mnemonic: E UN B SC (Every UNnecessary Bone SCan)

Duncan Chapter 4 (Risk), Page 92

Question 3

Sources of data for a clinical identification algorithm

Answer 3

1. Diagnosis in a medical record - highly reliable, but seldom available for actuarial work
2. Medical claims - one of the most common sources
3. Drug claims - the other most common source
4. Laboratory values
5. Self-reported data

Duncan Chapter 4 (Risk) - Page 92

Question 4

Definitions of sensitivity and specificity

Answer 4

When building clinical identification algorithms, the proper balance between sensitivity and specificity must be found

1. Sensitivity - the percentage of members correctly identified as having a condition (“true positives”)
2. Specificity - the percentage of members correctly identified as not having a condition (“true negatives”)

Note - One way to test for presence of false positives it to identify how many members re-qualify for the condition next year

Duncan Chapter 4 (risk) - Page 98

Question 5

External Sources of Clinical Identification Algorithms

Answer 5

1. HEDIS (from the NCQA) has algorithms for identifying some conditions (e.g., asthma, high blood pressure, and diabetes)
2. The Population Health Alliance - publishes a journal that evaluates many different types of intervention programs. Code sets for identification are frequently provided in these articles
3. CMS - Chronic Conditions Data Warehouse (CCW) provides researchers with Medicare and Medicaid data. A section of the CCW provides condition algorithms for more than 60 chronic or potentially disabling conditions
4. Quality reporting and improvement organizations - there are many of these organizations in the U.S.. Their publications can be a source of clinical algorithms
5. Grouper models - commercially-available models that identify member conditions and score them for relative risk and cost
6. Literature - articles will sometimes report the codes that are used for analysis

Mnemonic - HA PG CW QP GC LC

(HA PJ Can’t Win Quit Playing Go Cry LC?)

Duncan Chapter 4 (risk) - Page 103

Question 6

Major publishers of health care quality data measures

Answer 6

1. National Quality Forum (NQF) - has the lead responsibility in the US for determining which health care quality measures should be recognized as national standards. Operates under a three-part mission to improve the quality of healthcare by:
   a) Building consensus on national priorities and goals for performance improvement
   b) Endorsing national consensus standards for measuring and publicly reporting on performance
   c) Promoting the attainment of national goals through education and outreach programs
2. Agency for Healthcare Research and Quality - has developed a series of Quality Indicators (QIs) which use hospital data to highlight potential quality concerns. The QIs include inpatient, prevention, patient safety, and pediatric indicators.
3. Joint commission - The primary accrediting body for hospitals, nursing homes, and other care facilities
4. CMS - works with health care providers to develop measures of quality. Has the ability and funding to sponsor various quality initiatives
5. Hospital Quality Alliance - was formed to develop performance measures of hospital care. One of its products is the “Hospital Compare” website.
6. Measures applications Partnership (MAP) - a public-private partnership convened by the NQF to provide input on the selection of performance measures for public reporting and performance-based payment programs
7. American Medical Association Physician Consortium for Performance Improvement - a physician-led consortium focused on clinical quality improvement and patient safety

H CAN JAM (Health quality publishers CAN JAM!)

Duncan Chapter 4 (Risk), Page 108

Question 7

Reasons for using commercially-available grouper models

Answer 7

(Why these models may be preferred over building your own clinical identification algorithm)

1. Building algorithms from scratch requires a considerable amount of work
2. Models must be maintained to accommodate new codes, which requires even more work
3. Commercially-available models are accessible to many users. Providers and plans often require that payments be based on a model that is available for review and validation
4. CMS requires the use of a specific grouper model for risk adjustment in Medicare Advantage and ACA plans

(Note - TIA Chapter 5 hounds has printable list of grouper’s / table)

Duncan Chapter 5 (Risk), Page 111

Question 8

Principles for developing grouper models

Answer 8

(These principles guided the development of the Diagnostic Cost Groups. But they are also universal and continue to be promoted in other publications)

1. Diagnostic categories should be clinically (M)eaningful
2. Diagnostic categories should (P)redict medical expenditures
3. Diagnostic categories should have adequate sample sizes to permit stable (E)stimates
4. (H)ierarchies should be used to characterize the illness level within each disease process
5. Diagnostic classification should encourage (S)pecific coding
6. Diagnostic classification should not reward coding (P)roliferation
7. Providers should not be penalized for recording (A)dditional diagnoses
8. The classification system should be internally (C)onsistent
9. The diagnosis system should (A)ssign all codes (ICD-9/10)
10. (D)iscretionary diagnostic categories should be excluded

Mnemonic - MAP SHAPE C D (MAP to SHAPE Clinical Diagnoses)

Duncan Chapter 5 (Risk), Page 112

Question 9

Commercially-available diagnosis-based grouper models

Answer 9

1. Diagnostic Related Groups
   a) Relates type of patients treated to the cost incurred by hospital
   b) Used Extensively by CMS and some commercial payers to ensure consistent reimbursement of hospitals for patients with the same risk profile
2. Hierarchical Condition Categories (HCCs)
   a) CMS-HCCs - was developed as a health adjuster for Medicare Health plans
   b) HHS-HCCs - was developed for risk normalization for ACA plans
3. Clinical Risk Groups
   a) Similar to DRG, and can help identify clinically meaningful groups who require similar amounts and types of resources
   b) Classifies members of the population based on their burden of chronic medical conditions
4. Optum (Impact Pro) - is primarily used for predicting high-risk patients for care coordination
5. Chronic Illness and Disability Payment System
   a) developed for adjusting capitated payments for Medicaid beneficiaries
   b) CDPS+Rx - combines pharmacy
6. DxCG Intelligence
   a) uses patient-level information to profile the range and intensity of medical problems for a given population
   b) use to identify opportunities for early intervention
7. Symmetry - the Symmetry Episode grouper technology is the basis for Optum’s predictive and risk adjustment models
8. The Johns Hopkins Adjusted Clinical Groups System (ACG) -
   a) Patient demographics are merged with diagnoses and pharmacy information to produce a series of risk factors and scores.
   b) Primarily used for care management
9. Milliman Advanced Risk Adjusters (MARA)
   a) developed by actuaries and healthcare consultants to provide a set of risk adjustment models that predict risk at a more detailed level than was traditionally available. (Separate list for dimensions of this model)
   b) Prospective, predictive model to identify likely “next” group of high risk individuals
10. SCIO Prospective Financial Risk Model -
    a) an all-encounter model used to predict risk scores prospectively
    b) used for risk stratification, intervention program ROI, and case-mix adjustments
11. Risk- and Severity-adjustment methodologies for measuring inpatient quality care (Truven Health Analystics) - created to address mortality, complications, readmissions, and length of stay
12. Wakely Risk Assessment model - a transparent, high-performance, and open-code risk assessment model for a commercial population
13. Agency for Healthcare Research and Quality (AHRQ) Clinical Classifications Software (CCS) - a diagnosis and procedure categorization scheme

Duncan Chapter 5 (Risk), page 115

Question 10

Useful Dimensions of the MARA Model

Answer 10

1. Comprehensive set of risk scores by service category (6 scores per individual)
2. Inpatient and emergency room scores correlate strongly with the probability of admission and emergency room events
3. Individual condition profiles
4. Recency of care - identification of the most recent month of treatment of each condition
5. Persistency of Care - the number of instances of care for each condition
6. Risk Drivers - identifying the contribution of each medical condition
7. Chronic and non-chronic mapping of each condition group to accommodate easier cohort analysis
8. Identification of issues related to frailty

Duncan Chapter 5 (Risk), Page 124

Question 11

Commercially-available episode grouper models

Answer 11

Episode grouper models group all services that are associated with a particular diagnosis or procedure into a single group

1. Medicare Episode Grouper Model - developed by CMS for organizing administrative claims into information about resource use that can be used to support various program objectives
2. Episode Treatment Groups - a case-mix adjustment system that combines inpatient, ambulatory, and pharmaceutical claims to build a complete treatment episode from onset of symptoms until treatment is complete
3. Truven Medical Episode Groups - used by payers and providers to compare medical and surgical options and costs in the treatment of diseases of medical conditions

Duncan Chapter 5 (Risks), Page 129

Question 12

Core principles of episode grouper models

Answer 12

(These are the principles followed for the Truven Medical episodes groups model)

1. An episode of care considers all care for one medical condition for one patient
2. An episode should be described by the condition for which the patient was diagnosed, not the treatment the patient received
3. Different levels of severity within a condition should be accounted for by an episode grouper
4. Over time, a patient’s diagnosis may evolve, and the episode grouper should accommodate this within a single episode of care
5. An episode classification system should be clinically meaningful to providers
6. An episode classification system should be comprehensive, yet parsimonious and transparent

Duncan Chapter 5 (Risk), Page 131

Question 13

Types of Drug Grouper Models

Answer 13

1. Therapeutic class groupers - they group drugs into a hierarchy of therapeutic classes. Examples are American Hospital Formulary Service and Generic Product Identifier
2. Drug-based risk adjustment models - they infer the member’s diagnosis from the therapeutic class of drugs the member uses, and generate a relative risk score. Examples are Medicaid Rx, Pharmacy Risk Groups, and RxGroups

Duncan Chapter 5 (Risk), Page 132

Question 14

Diagnosis Related Groups (DRGs) -

Answer 14

• Health Care Finance Administration - DRG (HCFA-DRG) - used for medicare prospective payment system for hospitals (this is now renamed CMS)
• Refined-DRG (R-DRG) - introduces presence of absence of complications and co-morbidities (CC)
• Severity DRG ( S-DRG) - Refined DRGs to more adequately adjust for patient severity
• All Patient-DRG (AP-DRG) and All Patient Refined-DRG (APR_DRG) - Modification of HCFA-DRG that provides support for transplants, high-risk obstetric care, nutritional disorders, and pediatrics

Duncan Chapter 5 (Risk) (TIA Video)

Question 15

Common Features of Medicare Prospective Payment Systems

Answer 15

1. System of (A)verages - providers cannot expect to make a profit on each case, but efficient providers can make a reasonable return on average
2. Increased (C)omplexity - DRGs are more complicated than a system based on per Diem payments
3. (R)elative weights - associated with each patient group to reflect the average resources used by efficient providers
4. (C)onversion factor (base price) - the dollar amount for a unit of services. It is multiple by the relative weight to determine payment
5. (O)utliers - usually cases that require above-average resources and receive extra payments
6. (U)pdates - the conversion factor and relative weights are adjusted annually to reflect new technologies and changing practice patterns
7. (A)ccess and quality - policymakers monitor PPSs and survey patients to ensure that beneficiaries have adequate access to high quality care and that providers are compensated adequately

Mnemonic - COURAge And Complexity; why courage? lol

Duncan Chapter 6 (Risk), Page 141

Question 16

Challenges with patient classification systems based on coding systems

Answer 16

D - Need for new DRGs - due to new diseases and new procedures

I - ICD coding - some codes may not be sufficiently precise as diseases and procedures are refined

U - Upcoding - providers may be tempted to exaggerate a patient’s secondary diagnosis to get paid more

S - New Coding Systems - adopting the new ICD-10 systems will be a major challenge for hospitals and CMS

Mnemonic: IS D U (IS Difficult to Upgrade)

Duncan Chapter 6 (Risk), Page 142

Question 17

Summary of DxCG Grouping Levels

Answer 17

Aggregated Condition Categories (ACC)
Number - 31
Application - Population profiling, reporting

Related Condition Categories (RCC)
Number - 117
Application - Population profiling, reporting

Condition Categories (CC)
Number - 394
Application - Clinical screening, reporting

Hierarchical Condition Categories (HCC)
Number - 293
Application - Making predictions, clinical screening, reporting

DxGroups
Number - 1,010
Application - Clinical screening, reporting

ICD-9 and ICD-10 Diagnostic Codes
Number - 69,000+
Application - Coding and reimbursement

Duncan Chapter 6 (risk), TIA Video

Question 18

Two major types of DxCG Models

Answer 18

1. Concurrent - Used to reproduce actual historical costs
2. Prospective - Predicts what costs will be for a group in the future based on inherent conditions

Duncan Chapter 6 (risk), TIA Flash Card

Question 19

Goals of risk Adjustment for Arizona Medicaid Program

Answer 19

A - Align payment with the relative health risk of members at each health plan

A - Be Accurate and unbiased

• Accurate - there should be relatively high correlation between the projected cost of the population and the actual cost
• Unbiased - the methodology should not overcompensate for some risk factors at the expense of others

S - Be as Simple as possible while accomplishing other goals

B - Minimize the administrative Burden of developing and implementing the methodology

N - Be budget Neutral

Mnemonic - A B N S A (Arizona Budget Neutral State Adjustment)

Duncan Chapter 13 (risk), Page 278

Question 20

Methodology used to develop the Arizona Medicaid risk adjustment model

Answer 20

1. Model selected - Symmetry’s Episode Risk Groups (ERG) Model
   a) ERGs based on episode treatment groups
2. Type of data used - diagnosis codes and procedural information from medical data and national drug codes from pharmacy data
3. Data timing - three months of claim run-out was used
4. Eligibility groups - risk adjustment was applied to prospective, non-reconciled risk groups
5. Model Calibration - the model was re-calibrated by developing risk weights through a linear regression model based on Arizona Medicaid data, and then credibility weighting those rates with the model’s original risk weights.
6. Geographic issues - risk adjustment will take place at the geographical service area and risk group level
7. Individual approach - risk scores calculated during the experience period will follow the individual during the rating period. This will accurately reflect movement of individuals between health plans.
8. Risk factors are updated once per year
9. Risk factors for new members - members with at least 6 months of enrollment (“Long” cohort) during the experience period will be given a claims-based risk factor. Other members (“Short” cohort) will be given a risk factor that is the average of an age-gender factor and an adjusted plan factor
   a) Adjusted plan factor = (avg ERG risk score of long cohort / pure age-gender factor of long cohort) * Pure age-gender factor of short cohort
10. Phase-in - risk adjustment is being phased in such that only 80% of the 2009 rate is risk adjusted
11. Risk factors for newborns - a different approach is needed because newborns do not have prior year claims from which to develop condition-based risk scores. Claims of the prior cohort of newborns in the experience period are used to project newborn experience in the rating period

Duncan Chapter 13 (Risk), Page 278

Question 21

Formulas for calculating an MCO’s risk score for the Arizona Medicaid Program

Answer 21

1. Average ERG risks core for long cohort
   a) An unadjusted risk score is calculated as the sum over all risk factors and demographics of the risk weights multiplied by frequencies. Frequencies are the portion of the cohort with each risk factor or demographic
   b) For the Transitional Aid to Needy Families group, the final risk score equals the unadjusted risk score divided by a scaling factor
2. Total average risk score = % of members in long cohort * average ERG risk score for long cohort + % of members in short cohort * risk factor for short cohort
   a) Risk factor for short cohort = 50% * Adjusted plan factor for short cohrot (see formula in separate list) + 50% * pure age-gender factor of short cohort
3. The above formulas are calculated for the given MCO and for all MCOs in total. The MCO’s relative risk score = MCO total average risk score / average risk score for all MCO’s
4. Relative risk score with phase-in = 80% * relative risk score + 20% * 1.0000
5. A budget neutrality adjustment may also be applied to get the final relative risk score

Duncan Chapter 13 (risk), Page 285

Question 22

Formulas for calculating final capitation rates for the Arizona Medicaid program

Answer 22

1. Capitation rate to be risk adjusted = base capitation rate - bid risk contingency - bid admin cost - 2% premium tax
2. Risk-adjusted capitation rate (before retention) = Capitation rate to be risk adjusted * risk adjustment factor (this risk adjustment factor was referred to as the relative risk score in earlier formulas.
3. Final ris-adjusted capitation rate = risk-adjusted capitation rate (before retention) + bid contingency + bid admin cost + 2% premium tax

Duncan Chapter 13 (Risk), Page 287

Question 23

Common Hypotheses for the member selection patterns observed in Medicare Advantage plans versus traditional Medicare FFS

Answer 23

Those enrolling in Medicare Advantage plans have been observed to be materially healthier. Theories for why include:

1. Healthy enrollees are less reluctant to change benefit plans, so they are more likely to sign on with Medicare Advantage
2. Managed care organizations restrict access to certain network health care providers. Since less health Medicare enrollees generally have established provider relationships, they are more reluctant to leave traditional Medicare and risk losing access to their preferred providers

Duncan Chapter 14 (Risk), Page 291

Question 24

Methodology for calculating member risk scores for Medicare Advantage Part C

Answer 24

1. A member’s risk score is the sum of weights that reflect that member’s characteristics. This includes:
   a) An age/gender score
   b) A health condition score based on the coefficients attached to 79 different health hierarchical condition categories (HCCs)
2. A member may have multiple HCCs, and weights are included for each applicable HCC
3. There are also several weights that result from interactions between HCCs
4. weights are applied hierarchically. So if a member has multiple HCCs in the same hierarchy, only the weight for the most severe HCC is counted. The weights of the less-severe HCCs are “trumped”
5. Condition scores are prospective factors. Diagnoses in the prior year are used to predict Medicare health claim costs int he current year.
6. For members that are new to medicare, CMS provides only age/gender factors. These factors are higher than those for ongoing beneficiaries because the full responsibility for predicting future cost is assigned to only the age/gender factors.

Duncan Chapter 14 (Risk), Page 292

Question 25

Categories of Medicare Advantage and Part D Members

Answer 25

Members are split into these different categories for determining part C risk scores. Each category has its own age/gender and HCC factors

Categories for ongoing members:

1. Community, Non-dual eligible, Aged
2. Community, Non-dual eligible, Disabled
3. Community, Dual eligible - Full Benefits - Aged
4. Community, Dual eligible - Full Benefits - Disabled
5. Community, Dual Eligible - Partial Benefits - Aged
6. Community, Dual Eligible - Partial Benefits - Disabled
7. Institutional

Categories for new enrollees:

1. Non-Medicaid and not originally disabled
2. Medicaid and not originally disabled
3. Non-Medicaid and originally disabled
4. Medicaid and originally disabled

Categories for Part D Risk Adjustment

1. Aged, Low Income (LI)
2. Aged, Non Low Income (NLI)
3. Disabled, Low Income (LI)
4. Disabled, Non Low Income (NLI)

Duncan Chapter 14 (Risk), Page 293

Question 26

Experience items included in the Medicare Bid Pricing Tool (BPT)

Answer 26

The BPT is an excel workbook pricing form for Medicare Advantage plans to use. The following past experience items are projected forward two years from the base year to the contract year.

1. Average population risk score
2. Enrollment level (in member months)
3. Revenue
4. Claims
5. Non-benefit expense
6. Profit

Duncan Chapter 14 (Risk), Page 298

Question 27

CMS requirements for projected risk scores in the BPT

Answer 27

The formulas for projecting risk scores is established by CMS. But each Medicare Advantage Organization (MAO) is allowed to develop its own trend rate to use in that projection. The projected risk scores must:

1. Be based on the methodology for calculating the risk scores as discussed in the Rate (A)nnouncement
2. Be calculated using the (C)MS-HCC risk adjustment model
3. Reflect the (E)xpected risk score trend at the bid level
4. Be (A)ppropriate for expected population
5. Be adjusted for FFS (N)ormalization
6. Include the appropriate Medicare Advantage (C)oding adjustment factor

Two potions - ACE And Non Compliance (Mnemonic)

Duncan Chapter 14 (Risk), page 298

Question 28

Medicare vs ACA health Insurance Exchange Risk Adjustment

Answer 28

Medicare - Prospective Risk adjustment

ACA - Concurrent risk adjustment

Question 29

Reasons the ACA was enacted

Answer 29

(Or otherwise, goals, of the ACA)

1. Increase quality and affordability of health insurance
2. Lower the uninsured rate by expanding publinc and private insurance overage
3. Reduce the costs of healthcare for individuals and the government

Duncan Chapter 21 (Risk), Page 411

Question 30

Ways in which the ACA addresses the antiselection and instability created by guaranteed issue requirement

Answer 30

1. Subsidies are available for applicants with limited incomes
2. Employers are required to provide insurance. And all residents ineligible for employer coverage must purchase individual coverage, or pay a penalty
3. A risk adjustment mechanism transfers revenue from plans with relatively low-risk populations to plans with relatively high-risk populations. Risk corridors also transferred revenue from profitable to unprofitable plans

Duncan Chapter 21 (Risk), page 412

Question 31

Rating Factors allowed by the ACA

Answer 31

1. Age (limited to a 3:1 rate variation)
2. Geographic location
3. Family Size
4. Tobacco use (rates can be increased by 50% for tobacco users)

Duncan Chapter 21 (Risk), Page 415

Question 32

Reasons why the ACA uses a concurrent risk adjustment model

Answer 32

1. For the first year of the ACA, most exchange participants were expected to be previously uninsured. So no historical data was available to perform a prospective calculation
2. Prospective risk adjustments models are less accurate than concurrent models, as demonstrated in different Society of Actuaries comparative studies
3. The churn rate of members through the exchanges has been high, so even in later years many plans still will not have claims data on members

Duncan Chapter 21 (Risk), Page 415

Question 33

The ACA risk transfer formula

Answer 33

Ti = ( PLRS * IDF * GCF / Sum ( si * PLRS * IDF * GCF) - AV * ARF * IDF * GCF / Sum ( si * AV * ARF * IDF * GCF ) ) * P

1. Ti is the transfer amount per billable member month. A positive amount means the plan receives a payment, negative means plan makes a payment
2. IDF is the induced demand factor
3. GCF is the geographic cost factr
4. AV is the actuarial value
5. ARF is the average of the rating factors
6. si is the market share

P is the market-wide average premium

Duncan Chapter 21 (Risk), page 416

Question 34

Practical Issues with applying risk adjustment models

Answer 34

1. Risk Transfer models generally assume that risk and cost are correlated, so a 1% increase in risk is assumed to increase costs by 1%. But not all cost-risk relationships are linear. As a result, these models overcompensate some plans and under-compensate other plans
2. The Medicare Payment Advisory Commission (MedPAC) identified the following issues related to Medicare HCCs:
   a) Although the CMS-HCC risk adjusters map diagnosis codes to 189 HCCs, only 70 HCCs are actually used for irks-scoring. Some fairly prevalent medical conditions are not accounted for
   b) There is consdierablle variation within HCC’s in terms of patient severity and experience
   c) Certain Racial groups and income levels are likely to be higher consumers of healthcare, but this is not reflected in the model
   d) Because the model only uses one year of data for determining risk scores, for some chronic conditions the model under predicts since the patient doesn’t have a claim each year
   e) The standard model does not include a factor for the number of conditions. But MedPAC has found that this factor would lead to more accurate predictions.
3. Several issues exist in ACA risk adjustment (see separate lists for the Mass and national models)

Duncan Chapter 21 (Risk), pages 418 and 424)

Question 35

Problems with the Massachusetts risk adjustment model

Answer 35

1. Risk adjustment applies to the gross premium, not the cost of insurance or pure premium. So transfers include part of the expense margin
2. The model has been shown to be biased against zero-condition members, particularly at the younger ages
3. There is also a bias against limited network and other lower cost plans. Risk transfers have been observed to exceed net income for some of these plans
4. Risk adjustment operates at the state, rather than regional, level. This also creates a bias.

Duncan Chapter 21 (Risk), Page 426

Question 36

Issues and potential improvements for the national ACA risk adjustment model

Answer 36

1. Model is not accurate for adults with partial year enrollment (explains in greater detail in a list from GHS-120-17)
2. Lack of historical data - the ACA uses only one year of claims data in a concurrent model, which fails to properly reflect the risk chronic members who may not have a claim in some years
3. Only a fraction of members trigger conditions - this could be because a provider fails to code a condition or because the condition present is not mapped to an HCC
4. Because risks cores do not track costs well at the extreme, high-risk members may experience costs that are disproportionate to their risk scores.
5. Prospective vs concurrent models - sufficient data may now exist to move to a prospective model, but CMS has rejected making this change because it believes the concurrent model is best for this population
6. Market-share - insurers with small market shares and with premiums that are much different than the statewide average are likely to see revenue transfers that are unrelated to their own premiums

Duncan Chapter 21 (Risk), Page 428

Question 37

Key learnings from disease management programs that ACOs should apply to be successful

Answer 37

1. Any DM program, to be successful, needs to employ high quality data analytics, as close to real-time as possible
2. Medical records need to have the analytical sophistication and workflow capabilities to support the program. ACOs emphasize electronic medical or health records, but many of these are simply a repository of data and are not universally used by providers
3. Systems need to be aggregated before they can usefully support the ACO
4. The importance of economics
   a) Changing patient behavior in a way that produces a measurable financial outcome is a long and difficult undertaking
   b) Programs need to be focused on the patients who represent the greatest opportunity for cost reduction
5. The importance of planning and understanding the opportunity - economically successful programs must be focused. For example, they must identify what patients and what conditions will be managed

Duncan Chapter 22 (Risk), Page 434

Question 38

Structure of Medicare ACOs

Answer 38

1. An ACO is a network that is either physician-practice based or hospital based that shares responsibility for providing care to patients
2. The Medicare Shared Savings Program has two models of gainsharing
   a) One-sided - the ACO and CMS share 50/50 in any gains
   b) Two-sided - the ACO shares more of the gains, but is at risk for any losses
3. The ACO must meet certain requirements to be allowed to share savings with CMS
   a) The ACO must meet certain quality standards in the following domains: patient / caregiver experience, care coordination / patient safety, preventive health, and at-risk population
   b) Savings must surpass a hurdle rate, which ranges between 2% for the largest ACO’s and 4% for smaller ACOs
4. The ACO must manage all of the medical health care needs of at least 5,000 Medicare beneficiaries for at least three years
5. Patients do not enroll in the ACO. They are “attributed” to an ACO because they have received the plurality of their primary care from an ACO provider
   a) The patient is assigned to a PCP who is accountable for providing quality care, reducing utilization, and convincing the patient not to seek care outside the ACO provider network

Duncan Chapter 22 (Risk), Page 437

Question 39

Ways in which provider group-based ACOs are expected to generate savings

Answer 39

1. The practice should implement “Care coordination” to manage the care of the patients who need additional services
2. Access to integrated medical records and consistent management by physicians should reduce the need for tests
3. The ACO should develop a network of efficient providers and limit the use of less efficient providers
4. The focus on quality should result in fewer unnecessary services and better population health

Duncan Chapter 22 (Risk), Page 438

Question 40

Criteria for a beneficiary to be assigned to a participating ACO

Answer 40

1. The beneficiary must have a record of Medicare enrollment
2. The beneficiary must have at least one month of Part A and Part B enrollment, and cannot have any months of Part A only or Part B only enrollment
3. The beneficiary cannot have any months of Medicare group (private) health plan enrollment
4. The beneficiary may be assigned to only one Medicare shared savings initiative
5. The beneficiary must live in the United States or US territories and possessions.
6. The beneficiary must have a primary care service with a physician at the ACO
7. The beneficiary must receive the largest share of his or her primary care services from the particpating ACO

Duncan Chapter 22 (Risk), Page 440

Question 41

Steps in the process for CMS to assign beneficiaries to an ACO

Answer 41

Step 1 - the beneficiary is assigned to a participating ACO when:

a) The beneficiary has at least one primary care service furnished by a primary care practitioner, an
b) More primary care services (measured by Medicare allowed charges) are furnished by primary care practitioners at the participating ACO than from the same types of providers at any other ACO

Step 2 - for a beneficiary who has not received any primary care services from a primary care practitioner, the beneficiary is assigned to the participating ACO if:

a) The beneficiary received at least one primary care service from a specialist physician utilized in assignment at the participating ACO, and
b) More primary care services (measured by Medicare allowed charges) are furnished by specialist physicians utilized in assignment at a participating ACO than from any other ACO

Duncan Chapter 22 (Risk), Page 440

Question 42

Calculation of average per capita expenditure for ACOs

Answer 42

1. Expenditures are calculated for ACO-assigned beneficiaries separately for the following Medicare enrollment types:
   a) ESRD - eligibility for Medicare as a result of end state renal disease
   b) Disabled - eligibility for Medicare due to disability
   c) Aged/dual-eligible - eligibility for Medicare by age and eligible for Medicaid
   d) Aged/non-dual-eligible - eligible for Medicare by age and not eligible for Medicaid
2. Expenditures are defined as the total Medicare Parts A and B FFS payments from any provider for Shared Savings Program eligible months
3. Claims are assessed after three months of run-out. And a completion factor is applied by CMS
4. Average per capita expenditure = Sum (Claims(k) * t(k) ) / Sum ( tk ) with the summation from k = 1 to n
   a) The k values represent the different beneficiaries and t(k) is the exposure period of the kth beneficiary
   b) The calculation is done separately for each combination of Medicare enrollment type and benchmark and performance year

Duncan Chapter 22 (Risk), Page 441

Question 43

Risk Adjustment approaches for updating benchmarks to the performance years for Medicare ACOs

Answer 43

1. For newly assigned beneficiaries - the ACO’s CMS-HCC prospective risk scores are recalculated to adjust for changes in severity and case mix
2. For continuously assigned beneficiaries - for each performance year, the risk ratio is calculated as the ratio of the HCC score for that year relative to benchmark year 3. An overall risk ratio is calculated as a weighted average of the ratios for the different Medicare enrollment types.
   a) When the risk ratio is greater than one, demographic risk scores are used. This is done to negate some of the effect of diagnosis-driver increases in risk scores.
   b) When the risk ratio is less than one, HCC ratios are applied

Duncan Chapter 22 (Risk), Page 443

Question 44

Components of the ACA risk Adjustment Methodology

Answer 44

1. HHS-HCC risk adjustment model (HHS = Department of Health and Human Services; HCC = hierarchical conditional categories) - Uses an individual’s demographics and diagnoses to predict medical expense risk. A risk score is then calculated as a relative measure of how costly that individual is anticipated to be to the plan
2. Risk Transfer Formula - averages all individual risk scores in a covered plan, makes certain adjustments, and calculates the funds transferred between plans

GHS-119-17, Page 1

Question 45

Adaptions made to CMS-HCCs to develop HHS-HCCs

Answer 45

1. Prediction year - the CMS-HCC risk adjustment model is prospective, but the HHS-HCC risk adjustment model is concurrent
2. Population - the CMS-HCCs were developed using data from the elderly (age 65+) and disabled Medicare populations. The HHS-HCCs were modified to reflect medical conditions and cost patterns for commercial populations (under age 65)
3. Types of spending
   a) The CMS-HCCs are set up to predict non-drug medical spending, while the HHS-HCCs predict the sum of medical and drug spending
   b) The CMS-HCCs predict Medicare provider payments while the HHS-HCCs predict commercial insurance payments

GHS-119-17, Page 3

Question 46

Criteria used to determine which HCCs to use in the HHS risk adjustment model

Answer 46

1. Represent clinically-significant, well-defined, and costly medical conditions that are likely to be diagnosed, coded, and treated if they are present
2. Are not especially subject to discretionary diagnostic coding
3. Do not primarily represent poor quality or avoidable complication of medical care
4. Identify chronic, predictable, or other conditions that are subject to insurer risk selection, risk segmentation, or provider network selection., rather than random acute events that represent insurance risk

GHS-119-17, Page 4

Question 47

Data and methods used for developing the HHS-HCC risk adjustment model

Answer 47

1. Model Type - a concurrent model (instead of a prospective model) was chosen because no prior year information on health status existed for this population when the model was developed
2. Claims data from a large national proprietary database sourced form large employers and health plans was used to calibrate the model. Data was used only for enrollees who had coverage comparable to the essential health benefits under the ACA.
3. Expenditures - the model predicts expenditures for which plans are liable to create a plan liability risk score (PLRS)
4. Demographics and diagnoses - age ranges were created (e.g., five year intervals for adults). And only diagnosis codes from sources allowable for HHS risk adjustment are included.
5. Subpopulations - due to the clinical and cost differences in the adult, child, and infant populations, separate risk adjustment models were developed for each group. Separate models were also developed for reach cost sharing level (catastrophic, bronze, silver, gold, and platinum)
6. Model estimation - weighted least squares regression was used for determining model coefficients
7. Predicted plan liabilities - for each enrollee, a total predicted plan liability is calculated (see separate list)
8. Model evaluation
   a) The predictive accuracy at the individual level is measured by R-Squared
   b) The performance for subgroups is measured by the “predictive ratio,” which is the ratio of predicted to actual plan liabilities

SEED DEPT (Subpops, Estimation, Evaluation, Data, Demo/Diag, Expenditures, Pred.PlanLiab, ModelType)

GHS-119-17, Page 4

Question 48

Formulas for calculating HHS-HCC risk scores

Answer 48

1. Calculation of predicted plan liabilities for individuals - the total PLRS is the sum of the incremental predicted plan liabilities (coefficients) from the relevant model (based on the enrollee’s age and cost sharing level)
   a) for adults and children, this is the sum of the age/sex, HCC, and disease interaction coefficients
   b) For infants, this is the sum of the maturity/disease-severity category and additive sex coefficients
   c) Some individuals are eligible for reduced cost sharing. An induced demand factor is multiplied by the above sum to determine the final PLRS for them
2. Calculation of plan average PLRS - this is the plan’s weighted average of individual PLRSs, where the weights are enrollment months. All plan enrollees are counted in the numerator, but only billable plan enrollees (parents and the three oldest children) are counted in the denominator.

GHS-119-17, Page 23

Question 49

Options for improving the HHS-HCC risk adjustment methodology

Answer 49

1. Improve the accuracy of the model for partial year enrollees
   a) Length of enrollment could be included as a new indicator variable
   b) Or separate models could be produced for different enrollment period groups (months 1-4, 5-8, and 9-12)
   c) The second approach appears to predict more accurately, but it may present false precision when predicting costs for conditions with small sample sizes and it adds to the complexity of the risk adjustment methodology, which already includes separate models by age and metal level
2. use prescription drug utilization as a predictor in the model (see separate lists)
3. Pooling of high cost enrollees
4. Evaluating concurrent and prospective risk adjustment models

These last two were not described

GHS-120-17, page 35

Question 50

Benefits of adding prescription drug utilization to the HHS-HCC risk adjustment model

Answer 50

1. Imputing missing diagnoses - drug utilization data may capture the existence of some conditions that are missing in diagnoses entered on medical claims, particularly for chronic conditions. The model can assume an individual has a certain condition because that is the condition the drug treats
2. Severity indicator for a specific diagnosis - The presence of certain drugs can indicate the severity of illness for some HCCs
3. More timely, standardized data - drug data can be available more quickly than medical claims data, is often more complete, is often easier to access, and is more standardized because it does not vary with provider coding patterns
4. Mitigates the financial disincentive to prescribe expensive medications - a risk adjustment model that incorporates prescription drug utilization will compensate plans that cover high-cost medications, reducing the incentive for plans to restrict access to these medications

GHS-120-17, Page 40

Question 51

Concerns ab out adding prescription drug utilization to the HHS-HCC risk adjustment model

Answer 51

1. Risk adjustment (M)odels that use drug information are not as common as models based only on medical information, so they are not as well understood or accepted
2. (G)aming, perverse incentives, and discretionary prescribing
   a) Gaming occurs when a drug is prescribed in order to trigger a higher payment. Drug models are particularly susceptible to gaming because some relatively low-cost drugs are linked to high medical costs
   b) Financial incentives may inappropriately influence treatment decisions
3. Sensitivity of risk adjustment to (V)ariations in prescription drug utilization - many factors other than health status affect drug utilization, and risk adjustment based on drug utilization will reflect these factors (separate list of factors)
4. Added (A)dministrative burden (to calibrated and apply the model), operational complexity, and costs (due to data reporting requirements and frequent updates)
5. Availability of (O)utpatient drug data only - some drug models omit drugs provided in a hospital setting, which may make hospitalized patients appear to be less severely ill
6. (M)ultiple indications for most drugs - many drug classes are widely prescribed “off label” for indications that are not FDA-approved. So utilization of these drug classes does not always indicate the presence of a specific diagnosis.

MOM GAVe (us her concerns)

GHS-120-17, Page 41

Question 52

Factors other than health status that affect drug utilization

Answer 52

1. Plan and physician prescribing patterns
2. Cost sharing features
3. Drug utilization management features
4. Proclivities of providers for using drug versus non-drug treatments for a medical condition
5. The income level of enrollees

GHS-120-17, Page 42

Question 53

Criteria for evaluating hybrid risk adjustment models

Answer 53

“Hybrid” models are those that incorporate both diagnoses and prescription drugs

1. Clinical face validity - should be clinically valid in the relationship between the risk markers (diagnoses and drugs) and health care expenditures, and in the relationship between drugs and associated diagnoses
2. Empirical / predictive accuracy - drugs added to the model should increase the mode’s accuracy in predicting health expenditures
3. Incentives for prescription drug utilization - adding drugs should be done in a way that minimized incentives for over-prescription of drugs to maximize risk transfers, but does not discourage needed drugs
4. Sensitivity to variations in prescription drug utilization - should incorporate variations in drug utilization that measures differences in enrollee health status, not variation due to other factors
5. Incentives for diagnosis reporting - accurate and complete diagnosis reporting should not be discouraged by reduced predicted expenditures when additional diagnoses are appropriately reported

GHS-120-17, Page 44

Question 54

Approaches for adding prescription drug utilization to a risk adjustment model

Answer 54

1. Statistical predicative power approach - drug classes are included in the model on purely statistical grounds
   a) Advantage - this approach allows for a linkage between a drug and poor health in general
   b) Disadvantage - by omitting clinical considerations it makes interpretation of model coefficients difficult and leads to less clinical face validity
2. Conceptual approaches for adding drug utilization to a diagnosis model to create a hybrid model (the first two were identified as the “two main approaches”)
   a) Imputation - using drug data to impute missing diagnoses. The predicted incremental cost is the same regardless of how a health condition is identified, whether by a drug indicator only, a diagnosis indicator only, or both indicators
   b) Severity - using drug data as a severity indicator for a specific diagnosis. Only if the drug class and a specific diagnosis are both present will the model predict incremental costs beyond the diagnosis alone.
   c) Rx Dominant - individuals taking a drug are assumed to be more severely ill (have higher projected costs) than individuals not taking the drug who only have the associated diagnosis marker
   d) Flexible, generalized empirical framework - each drug-diagnosis pair enters the model with three indicator variables: A diagnosis indicator, a drug class indicator, and an interaction indicator. Each indicator has a coefficient that predicts the incremental costs for that indicator

GHS-120-17, Page 45

Question 55

Criteria for selecting drug-diagnosis pairs for a hybrid model

Answer 55

1. Select drugs with patterns of non-discretionary prescribing
2. Avoid drugs where there are incentives for over-prescribing
3. Avoid drugs where there are variations in prescribing across providers, practices, and areas
4. Carefully consider selection of high-cost drugs. In some cases, including the drug in the model may reduce the incentives for insurers to strive for greater efficiency
5. Avoid drugs indicated for multiple diagnoses
6. Avoid drugs indicated for diagnoses not included in the HHS-HCC model
7. Carefully consider selection of drugs in an area exhibiting a rapid rate of technological change (which could make cost predictions inaccurate when based on previous years of data)

GHS-120-17, Page 49

Question 56

HHS considerations when selecting drug-diagnosis pairs to include in a hybrid HHS-HCC model

Answer 56

1. Empirical considerations - a wide range of exploratory data analysis was performed to determine pairs to consider. Then stepwise regression was used to dtermine which drug classes added the most predictive power to the existing HHS-HCC model.
2. Clinical considerations - doctors and pharmacists were consulted to provide deeper insights into the medical links between health conditions and the drug groups being considered and to identify the potential for gaming for each drug being considered
3. Additional considerations
   a) Imposing model restrictions based 2on day’s supply or number of prescriptions in order to trigger a drug indication
   b) Whether to split certain drug classes or restrict a drug-diagnosis interaction to certain drugs within a class
   c) HHS examined different models that include imputation-only versus imputation and severity approaches
   d) Prophylactic (preventive) use of drugs - drugs are sometimes used in persons at risk of a disease but who do not actually have the disease
   e) Multiple indications for drugs - drug classes are often indicated for multiple diagnoses

GHS-120-17, Page 50

Question 57

Considerations when selecting risk characteristics to use in a risk classification system

Answer 57

1. (R)elationship between the risk characteristics and expected outcomes. Rates are considered equitable for a given risk characteristics if differences in rates reflect material differences in expected cost
2. (C)ausality - the risk characteristic should be related to expected outcomes, but it is not necessary to establish a cause and effect relationship
3. (O)bjectivity - select risk characteristics that are capable of being objectively determined
4. (P)racticality - reflect the tradeoffs between practical and other relevant considerations
5. (A)pplicable law - consider whether the law limits the choice of risk characteristics
6. (I)ndustry practices - consider usual and customary risk classification practices for the given situation
7. (B)usiness practices - consider limitations created by business practices for the given situation

BLIP OR C? (BLIP OR Causality?

ASOP #12, Page 4

Question 58

Considerations when establishing risk classes

Answer 58

1. intended use - select a risk classification system that is appropriate for the intended use
2. Actuarial considerations
   a) Adverse selection - may occur if the variation in expected outcomes within a risk class is too great
   b) Credibility - risk classes should be large enough for expected outcomes to be credible
   c) Practicality - must balance the conflicting objectives of accuracy and efficiency
3. Other considerations - should comply with applicable law, consider industry practices, and consider limitations created by business practices
4. Reasonableness of results from using the risk classes

ASOP #12, Page 3

Question 59

Considerations when selecting a risk adjustment model

Answer 59

1. intended (U)se - consider the degree to which the model was designed to estimate what the actuary is trying to measure
2. (I)mpact on program - consider whether the risk adjustment system may cause changes in behavior because of underlying incentives
3. Model (V)ersion - if a new version of a previously-utilized model is used, consider the materiality of changes to the model
4. (P)opulation and program - consider if the population and program to which the model is being applied and consistent with those used to develop the model
5. (T)iming of data collection, measurement, and estimation - consider the impact of timing differences between when the model is developed and when it is applied
6. (T)ransparency - consider whether the model provides an appropriate level of transparency for the intended use
7. Predictive (A)bility - consider the predictive ability of the model and the characteristics of the various common predictive performance measures
8. Reliance on (E)xperts - consider whether the individuals incorporating their specialized knowledge into the model are experts in risk adjustment
9. (P)ractical considerations - consider practical limitations, such as the cost of the model, the actuary’s familiarity with the model, and its availability

T PAVE IT UP (To PAVE IT UP)

ASOP #45, Page 3