This guide develops the actuarial skills behind provider reimbursement, intervention evaluation, U.S. health risk adjustment, medical data, social determinants, and predictive analytics. Twenty original written exercises emphasize calculations, defensible comparisons, and interpretation; sixty concepts explain mechanisms and common errors. It follows the supplied November 2026/March 2027/July 2027 syllabus, the next-administration edition as of October 16, 2026. Written exercises are a learning format; the supplied documents do not verify the examination's response format. Allocations are 5/5/4/3/2/1 questions and 15/15/12/9/6/3 concepts. Citations establish syllabus relevance, not independent verification of every technical detail. Contract examples are illustrative; this guide makes no claims about later syllabus versions.
What this guide covers
Study GH 301 with 20 original actuarial exercises and 60 concepts covering its six syllabus domains.
Provider Contracting and Reimbursement
1. Allowed amounts and fee-for-service liability
Fee-for-service reimbursement pays for separately identified covered services under an agreed fee schedule or other pricing rule. The allowed amount is the contractual reimbursement base; billed charges can differ substantially. Deductibles, coinsurance, and copayments allocate that base between the payer and member according to benefit terms. Provider payment, plan liability, and total allowed spending therefore answer different questions. Analyze units, modifiers, covered services, and contractual discounts before comparing prices. Increased service volume or intensity can raise total spending even when the negotiated price per unit remains unchanged.
Worked example: A specialist bills $900, but the contract allows $600. With the deductible satisfied and 25% coinsurance, the member pays $150 and the plan pays $450. A second identical service doubles allowed spending even though the unit price stays $600.
Common mistake: Comparing billed charges across providers can suggest price differences that disappear when the actual negotiated allowed amounts are examined.
Official syllabus reference: [2]
2. Per diem payment and length-of-stay incentives
Per diem reimbursement assigns payment to each covered inpatient day, sometimes with different rates by service, day, or acuity. Total reimbursement depends on both admissions and covered length of stay. An extra day generates additional payment, but profitability depends on its incremental cost and any authorization requirements. A single average daily rate can hide front-loaded costs or specialized units. Evaluate admission frequency, length of stay, transfers, readmissions, and outcomes together. Contract definitions must establish which days count and how observation, discharge days, and exceptional services are treated.
Worked example: A contract pays $1,100 per covered day for a four-day admission, producing $4,400. A fifth covered day adds $1,100 of revenue. Whether that day increases margin depends on the hospital's additional resource cost and the contract's coverage conditions.
Common mistake: A lower per diem rate does not guarantee lower episode spending when admissions or covered lengths of stay increase.
Official syllabus reference: [2]
3. DRG weights, case mix, and outlier provisions
A diagnosis-related group classifies an inpatient case for a specified grouping and payment system. A simplified payment multiplies a base rate by the assigned relative weight, with additional adjustments where applicable. The case-mix index is the average relative weight across admissions, using admission counts as weights. Fixed case payment shifts some within-admission resource risk to the hospital, while outlier provisions can limit extreme losses. Changes in case mix may reflect clinical severity, coding, or grouping rules. Comparisons require consistent groupers, versions, service boundaries, and payment adjustments.
Worked example: A hospital has 60 admissions at weight 1.0 and 40 at weight 1.5. Its case-mix index is 1.20. At an illustrative $6,000 base rate, average payment is $7,200 before any outlier, geographic, or other contractual adjustments.
Common mistake: An increase in average DRG weight does not independently demonstrate that patients became clinically sicker or that resource use increased proportionately.
Official syllabus reference: [2]
4. Capitation exposure and service carve-outs
Capitation pays a specified amount per covered member-period for a defined set of services. Revenue depends primarily on attributed or enrolled exposure, while financial results depend on the cost of delivering covered care. Service carve-outs leave designated costs outside the capitation arrangement, changing both transferred risk and incentives. Member eligibility, retroactive enrollment changes, referral responsibilities, and risk adjustment affect settlement. Compare capitation rates only after aligning covered services and population morbidity. A low rate with narrow coverage can transfer less financial risk than a higher rate covering comprehensive services.
Worked example: A physician group receives $45 PMPM for 8,000 member-months, yielding $360,000. If specialty drugs are carved out, those drug claims remain outside the group's cost responsibility. Adding them later requires an explicit assessment of expected cost and volatility.
Common mistake: Treating every claim for capitated members as the provider's responsibility ignores contractual exclusions and can materially misstate transferred financial risk.
Official syllabus reference: [2]
5. Risk corridors versus specific stop-loss
A risk corridor shares aggregate financial results when spending or losses move beyond specified contract boundaries. Specific stop-loss limits responsibility for exceptionally costly individual members or episodes according to defined attachment and reimbursement terms. These protections address different sources of volatility: overall experience versus concentrated large losses. Evaluate attachment points, reimbursement percentages, exclusions, maximum recoveries, and settlement order. Their effects can overlap, so contract interpretation must prevent duplicate adjustments. Neither arrangement removes all provider risk, and both can alter incentives for managing costs near the protected portions of the distribution.
Worked example: A group retains ordinary claims but receives 80% reimbursement for each member's covered annual claims above an illustrative $100,000 attachment. A separate corridor shares aggregate losses beyond the annual target. Recoveries must enter the aggregate settlement in the contractually specified order.
Common mistake: Calling an aggregate loss-sharing corridor individual stop-loss obscures which claims trigger protection and can produce an incorrect reconciliation.
Official syllabus reference: [2]
6. Episode boundaries and bundled-payment attribution
A bundled-payment episode groups services around a clinical event or condition within a specified period. Its financial meaning depends on the trigger, eligible population, included providers, service exclusions, and episode end. Attribution determines which organization is responsible when several providers participate. Overlapping episodes require precedence rules to prevent double counting. Complication and readmission provisions influence incentives for coordination and service shifting. Before estimating savings, construct comparable historical episodes using the proposed rules and test how changes in boundaries alter expected cost, variability, and provider accountability.
Worked example: A surgical bundle includes the admission and 60 post-discharge days but excludes unrelated trauma. A readmission for a surgical complication is included. Applying these boundaries to historical claims produces a different target than simply summing the initial hospital bill.
Common mistake: Using the hospital admission alone to price a broader episode omits post-acute services and can understate the provider's assumed cost responsibility.
Official syllabus reference: [2]
7. Prospective bundles and retrospective reconciliation
A prospective bundle makes a defined episode payment that must support the included providers and services. A retrospective bundle generally preserves underlying claims payments and later compares eligible episode spending with a target. Reconciliation rules determine sharing, repayment, quality adjustments, and risk protection. Cash-flow responsibility and administrative requirements differ even when economic targets are similar. Separate total episode spending from the participant's reconciliation amount and operating margin. Targets and actual claims must use consistent episode definitions, risk adjustments, completion assumptions, and exclusions before a financial difference is meaningful.
Worked example: Eligible claims total $970,000 against a $1 million target. A retrospective agreement awards the participant 50% of the $30,000 difference, producing $15,000. That amount is an additional settlement, not a replacement for the claims already paid to providers.
Common mistake: Adding the entire target to underlying claims payments in a retrospective bundle double counts the reimbursement base and greatly overstates provider revenue.
Official syllabus reference: [2]
8. ACO attribution and benchmark comparability
An accountable care organization's financial evaluation compares spending for an attributed population with a defined benchmark. Attribution can depend on utilization patterns or prospective assignment under program-specific rules. Population turnover, historical spending, regional comparisons, trends, and risk adjustments can all affect benchmark comparability. A spending difference can arise without an equivalent change in care delivery when assignment or benchmark construction changes. Evaluate member movement, continuous eligibility, service scope, and quality alongside financial results. Do not transfer one ACO program's attribution or benchmark rules to another without checking its governing specifications.
Worked example: An ACO's spending falls after many high-cost members leave its attributed population. Recalculating comparable exposure and morbidity may explain much of the decline. The analyst should distinguish that compositional effect from reductions achieved among otherwise comparable members receiving coordinated care.
Common mistake: Assuming every dollar below an ACO benchmark was caused by care coordination overlooks attribution changes and the benchmark's construction.
Official syllabus reference: [2]
9. Shared-savings thresholds, caps, and quality gates
Shared-savings contracts translate a spending difference into a provider settlement through separate qualification and payment rules. A minimum savings requirement may screen out small differences; the sharing percentage determines participation in eligible savings; a cap limits payment. Quality gates or adjustments condition financial rewards on measured performance. Some contracts also share losses, creating asymmetric or two-sided exposure. Read the order of operations carefully, especially whether sharing applies to all savings or only amounts beyond a threshold. These are contract-specific provisions rather than universal features of value-based reimbursement.
Worked example: An illustrative contract recognizes $2 million of eligible savings and shares 40%, subject to a $700,000 cap. The preliminary $800,000 reward becomes $700,000 if the quality requirement is met. A failed quality gate can change the payment independently of spending.
Common mistake: Subtracting a qualification threshold from savings without checking the contract can understate payment when the threshold merely determines eligibility.
Official syllabus reference: [2]
10. ACO REACH performance-year rule sensitivity
ACO REACH analysis requires separating historical performance-year rules from assumptions for a different performance year. Benchmark trends, risk-adjustment treatment, reconciliation provisions, and model changes can affect financial results independently of operational improvement. The assigned syllabus includes separate PY2025 and PY2026 readings, making version control essential. Build a bridge that holds population and spending constant while changing one rule at a time. Then analyze operational changes separately. The supplied outline establishes these readings' relevance but does not supply their detailed formulas, so exact year-specific provisions cannot be inferred from their titles.
Worked example: A finance team models identical members and claims under two hypothetical rule sets. The benchmark changes while actual spending remains fixed, altering projected settlement. That bridge quantifies a rule effect; it does not establish that the care model became more effective.
Common mistake: Carrying a favorable prior-year reconciliation effect into a later performance year without checking the applicable rules can create unsupported earnings expectations.
Official syllabus reference: [2]
11. Episode-based physician cost attribution
Episode-based physician profiling assigns condition-related spending to a physician using defined attribution rules. The resulting profile may include services ordered, delivered by other providers, or influenced by referral networks. Compare episodes within clinically comparable categories and adjust for measurable severity before interpreting cost differences. Attribution to a physician does not establish that every included cost was under that physician's control. Episode completeness, overlapping conditions, referral patterns, and selection into treatment can affect results. Use the profile to investigate variation, with explicit accountability boundaries, rather than treating attribution as a causal finding.
Worked example: Two physicians manage similar orthopedic episodes, but one sends patients to a higher-priced imaging facility. Episode costs differ even with identical imaging frequency. Separating unit prices, utilization, and clinical severity clarifies whether the profile reflects referral prices or treatment intensity.
Common mistake: Assigning all episode spending to one physician and calling it controllable cost confuses a measurement convention with actual decision authority.
Official syllabus reference: [2]
12. Reliability shrinkage and profile misclassification
A provider profile combines a performance signal with statistical noise. Reliability measures how much observed variation is attributable to persistent differences under the model rather than sampling variability. Credibility-style shrinkage moves less reliable estimates toward an appropriate reference mean. This reduces extreme rankings driven by small samples, although it can also attenuate genuine differences. Classification decisions require uncertainty and error consequences, not just adjusted point estimates. Reliability depends on outcome variation, case mix, sample structure, and model assumptions; episode count alone does not uniquely determine whether a profile supports tier assignment.
Worked example: An observed cost ratio of 0.70 with reliability 0.20 becomes 0.20 × 0.70 + 0.80 × 1.00 = 0.94 under a simple shrinkage rule. The striking raw result contributes limited evidence because most weight remains on the peer mean.
Common mistake: Treating every observed low-cost provider as truly efficient ignores sampling noise and can systematically misclassify providers with limited or unusually variable experience.
Official syllabus reference: [2]
13. Network tiering and member response
Network tiering steers members through differences in cost sharing, provider designation, or other incentives. Its financial effect depends on price differences, quality, accessibility, member switching, and provider responses. A tier based on low unit prices can produce different results from one based on risk-adjusted episode cost. Evaluate whether members can realistically access preferred providers and whether migration creates capacity constraints. Savings estimates should distinguish the redistribution of existing utilization from changes in total utilization. Classification reliability and transparent measurement rules matter because tier placement affects both provider revenue and member choices.
Worked example: A plan expects 500 visits to shift from providers allowing $180 per visit to providers allowing $150. The initial price-only opportunity is $15,000. Travel barriers, appointment availability, and any additional visits could change the actual spending and access results.
Common mistake: Multiplying the full population's claims by a preferred-tier discount assumes complete switching and overlooks access, capacity, and differences in service use.
Official syllabus reference: [2]
14. Fee-schedule escalation and lesser-of clauses
Contract pricing may combine a fee schedule, annual escalators, billed-charge limits, and effective-date rules. A lesser-of clause can make the payable base the lower of the scheduled allowance and a specified charge amount. An escalator therefore need not produce the same percentage increase in actual reimbursement. Analyze claims under the complete pricing sequence, including units, modifiers, amendments, and service dates. Separate changes in contractual unit prices from utilization and coding changes. A contract's stated percentage increase is an input to repricing, not automatically the measured increase in total paid or allowed claims.
Worked example: A contract pays the lesser of billed charges and its fee schedule. The scheduled fee rises from $100 to $105, but a provider bills $102. The allowed amount becomes $102, so realized reimbursement rises 2% rather than the schedule's 5%.
Common mistake: Applying a fee-schedule escalator directly to all historical allowed claims ignores binding charge limits and other terms that alter realized payment.
Official syllabus reference: [2]
15. Surprise-billing protection and payment disputes
Analysis of U.S. surprise-billing protections distinguishes the patient's cost-sharing obligation from the final payment negotiated or determined between a provider and payer. For services within applicable protections, patient liability generally follows specified in-network treatment rather than unrestricted balance billing. Provider payment disputes can follow separate processes whose applicability depends on service, coverage, jurisdiction, and current rules. The qualifying payment amount can play a defined role, but it should not be treated as interchangeable with billed charges or every final settlement. The outline's interview titles do not establish detailed legal rules.
Worked example: A protected emergency claim has a disputed provider payment. The analyst separately calculates patient cost sharing under the applicable rules and models the payer-provider settlement. A higher final settlement should not automatically be passed through as additional patient balance billing.
Common mistake: Using a provider's billed charge as the universal patient-liability base confuses the payment dispute with the separate protections governing member cost sharing.
Official syllabus reference: [2]
Disease Management
16. Opportunity analysis and the actionable-risk funnel
Care-management opportunity depends on potentially avoidable spending, reachable members, intervention effectiveness, and implementation cost. High predicted spending alone does not imply that spending is modifiable. An opportunity funnel starts with eligible exposure and applies distinct assumptions for engagement, preventable events, and incremental intervention effect. Avoid counting the same effect both as preventability and effectiveness. Segment opportunities by clinical pathway and program capability rather than treating every high-cost member alike. The resulting estimate is a planning scenario, while realized savings require a credible comparison and complete accounting of resources used.
Worked example: A program identifies 1,000 eligible members, expects 40% engagement, and assumes engagement prevents 0.05 admissions per person annually. At $15,000 per admission, the gross opportunity is $300,000 before program costs, assuming those effects are incremental and nonoverlapping.
Common mistake: Multiplying all high-risk members' spending by a broad reduction percentage overstates opportunity when much spending is necessary or members cannot be reached.
Official syllabus reference: [2]
17. Net ROI, benefit-cost ratios, and evaluation perspective
Net ROI divides incremental net financial benefit by incremental program cost; a benefit-cost ratio divides gross benefit by that cost. Both require explicit definitions because reports use inconsistent terminology. Include relevant implementation, staffing, vendor, and administrative costs without double counting. Match benefits and costs to the same population, period, and perspective. Payer medical savings differ from provider margin or societal value, and costs may move between organizations. A positive arithmetic ROI does not validate the counterfactual savings estimate. For multiyear programs, timing and discounting can materially change the result.
Worked example: A program estimates $500,000 of gross payer savings and costs $400,000. Net benefit is $100,000, net ROI is 25%, and the benefit-cost ratio is 1.25. A provider may experience a different result if reduced utilization also reduces its revenue.
Common mistake: Reporting a benefit-cost ratio of 1.25 as 125% net ROI confuses gross benefit per dollar with profit after recovering program cost.
Official syllabus reference: [2]
18. Actuarially adjusted historical-control construction
An actuarially adjusted historical control projects a baseline population's experience to the evaluation period as an estimate of what would have occurred without intervention. Adjustments may address trend, benefits, population risk, exposure, and other relevant changes. Each factor should have an explicit rationale and avoid duplicating another factor's effect. Compare the resulting expected cost with actual cost using consistent services and claims completion. The method's validity depends on the counterfactual projection, not merely on accurate arithmetic. Sensitivity testing should expose how plausible alternative trends or risk adjustments change estimated savings.
Worked example: Baseline cost is $700 PMPM, external trend is 8%, and a separately supported benefit factor is 0.98. Expected cost is $740.88 PMPM. Actual cost of $720 suggests $20.88 PMPM savings, conditional on the projection accurately representing untreated experience.
Common mistake: Applying both a fully benefit-adjusted trend and an additional adjustment for the same benefit change double counts an influence on expected spending.
Official syllabus reference: [2]
19. Difference-in-differences and parallel trends
Difference-in-differences subtracts a comparison population's outcome change from the intervention population's change. Its central identifying assumption is that the groups would have followed parallel untreated trends on the chosen outcome scale. Similar baseline levels are neither necessary nor sufficient. Check pre-intervention trajectories, enrollment changes, concurrent policies, and differential outcome measurement. An additive rate analysis and a multiplicative trend analysis encode different counterfactuals and can produce different estimates. A negative spending difference can support estimated savings under the assumptions, but the design does not automatically eliminate selection or time-varying confounding.
Worked example: Intervention cost falls from $900 to $850 PMPM while comparison cost rises from $600 to $620. The additive estimate is −$50 − $20 = −$70 PMPM. Its interpretation depends on whether the intervention group would otherwise have increased by $20.
Common mistake: Choosing a comparison group solely because its baseline spending matches the intervention group's ignores the untreated trend assumption needed for causal interpretation.
Official syllabus reference: [2]
20. Regression to the mean after high-cost selection
Selecting members because of unusually high recent spending captures both persistent risk and temporary random or episodic influences. On average, some temporary influences subside, so later spending can decline without an intervention. This regression to the mean is especially important for programs triggered by hospitalizations or extreme claims. A historical baseline alone cannot distinguish natural reversion from treatment benefit. Use a comparison selected by analogous criteria, longer baseline information, or a suitable design that accounts for the selection mechanism. Prediction of future spending and evaluation of incremental savings remain separate tasks.
Worked example: A plan enrolls members immediately after an expensive hospitalization. Their next quarter costs less even when no service changes occur. Comparing them with similarly selected eligible members receiving usual care helps distinguish ordinary post-event decline from the program's additional effect.
Common mistake: Calling every post-enrollment reduction a program saving ignores the temporary events that caused members to qualify for enrollment in the first place.
Official syllabus reference: [2]
21. Propensity scores and baseline treatment assignment
A propensity score estimates the probability of receiving an intervention conditional on observed pre-intervention characteristics. Matching, stratification, or weighting can use this score to improve comparability of measured baseline covariates. Covariates should reflect the treatment-assignment and outcome processes without conditioning on variables caused by treatment. A strong participation prediction model is not automatically a strong adjustment design; covariate balance is the operational target. Identification also requires assumptions about unmeasured confounding and overlap. The score does not measure expected clinical benefit, and equal scores do not prove that individual members are clinically identical.
Worked example: A coaching program attracts younger members with more prior office visits. An analyst estimates participation propensity from baseline age, diagnoses, and utilization, then matches participants to controls. Post-match comparisons must verify those characteristics are balanced rather than relying only on model discrimination.
Common mistake: Adding post-program utilization to the propensity model can condition on a consequence of treatment and distort the intended causal comparison.
Official syllabus reference: [2]
22. Common support, weighting, and covariate balance
Common support means the analysis contains plausible treated and untreated observations across the covariate profiles being compared. Extreme propensity scores can produce unstable inverse-probability weights or reveal that some participants lack suitable controls. Restriction, trimming, or alternative weighting can improve stability but change the target population or estimand. Assess balance for important covariates after adjustment, using measures such as standardized mean differences and distributional checks. No single universal cutoff establishes validity. Report retained exposure and effective information, because a large raw sample can still support a fragile weighted comparison.
Worked example: A participant with estimated treatment probability 0.99 has few comparable untreated members. Keeping an extremely weighted control may make the estimate depend on one person. Restricting the analysis can improve stability while limiting conclusions to members with genuine comparison support.
Common mistake: Assuming matching succeeded because average propensity scores agree overlooks remaining imbalance in individual covariates and instability from extreme weights.
Official syllabus reference: [2]
23. Chronic-cohort migration and status timing
Chronic and non-chronic trends depend on when disease status is assigned and whether cohorts remain fixed. A retrospective definition using future diagnoses can move high-cost members between categories and introduce information unavailable at baseline. A fixed baseline cohort answers a different question from annually reclassified populations. Newly diagnosed members may carry distinctive costs around diagnosis, affecting both group averages. Specify the lookback period, qualifying evidence, enrollment requirements, and treatment of incident cases. Decompose membership migration before attributing subgroup trends to disease management or changes in the cost of established chronic disease.
Worked example: A member classified as non-chronic in Year 1 receives a diabetes diagnosis in Year 2. Annual reclassification moves that member into the chronic group, while a fixed baseline analysis retains the original category. The two approaches legitimately produce different trend measures.
Common mistake: Comparing annually reclassified groups as though they contain the same members can misinterpret cohort migration as a change in treatment cost.
Official syllabus reference: [2]
24. Direct standardization of chronic and non-chronic trends
Direct standardization applies fixed population weights to subgroup-specific outcomes from each period. It separates within-group cost changes from changes in the proportion of members assigned to each group. The standardized result answers what aggregate cost would have been under a common composition; it is not the population's actual expenditure. Choose a reference distribution appropriate to the comparison and retain consistent subgroup definitions. Residual severity changes within groups can still affect results. More detailed risk categories may improve adjustment but can reduce credibility and increase dependence on coding and model assumptions.
Worked example: At fixed weights of 25% chronic and 75% non-chronic, costs rising from $1,200 and $200 to $1,320 and $220 produce a 10% standardized increase. Actual aggregate trend can be higher if the chronic membership share also rises during the year.
Common mistake: Treating standardized cost as the actual budget requirement ignores the current population mix that determines the dollars the plan must pay.
Official syllabus reference: [2]
25. Price, utilization, and service-mix decomposition
Medical spending reflects service frequency, unit prices, and the mix or intensity of services. A program can reduce admissions while total spending rises because prices increase or remaining admissions become more complex. Decomposition requires stable units and an explicit treatment of interactions among components. Reprice services at common prices to examine utilization and mix separately from negotiated payment changes. Clinical substitution can move spending between inpatient, outpatient, pharmacy, and community services. Evaluate the full relevant scope before interpreting one category's reduction as net program savings or a deterioration in care.
Worked example: Admissions decline from 100 to 90, while average payment rises from $10,000 to $12,000. Spending increases from $1 million to $1.08 million. The admission reduction is real, but a price-and-mix analysis is needed before judging total financial performance.
Common mistake: Equating fewer inpatient admissions with equal percentage savings ignores unit-price changes, case severity, and spending that shifts into other settings.
Official syllabus reference: [2]
26. Exposure denominators and recurrent-event rates
Utilization rates divide event counts by the exposure during which events could occur. Member-months or member-years accommodate enrollment periods that differ across members. Recurrent-event rates count multiple admissions, while the proportion of members admitted counts each person once; these measures answer different questions. Align enrollment, observation windows, service dates, and eligibility rules. A program that changes survival or retention can change observed exposure, complicating comparisons. Translate rate differences into event counts using the appropriate intervention exposure, and retain units throughout calculations to avoid confusing per-member, per-thousand, and annualized results.
Worked example: Thirty admissions occur over 150 member-years, yielding 200 admissions per 1,000 member-years. If only 20 members had any admission, the admitted-member proportion requires a separate eligible-person denominator. It cannot be inferred by treating all 30 events as distinct people.
Common mistake: Dividing events by year-end enrollment can distort rates when members entered or left coverage throughout the measurement period.
Official syllabus reference: [2]
27. Intention-to-treat and engagement-based effects
An intention-to-treat analysis compares groups according to original randomized assignment, preserving the assignment's protection against baseline confounding. It estimates the effect of offering or assigning the intervention under observed uptake. Comparing only engaged participants with controls generally loses that protection because engagement can reflect motivation, access, or health. A treatment-received effect requires additional methods and assumptions. For program planning, the offer effect may be especially relevant because real implementation also includes nonparticipation. Distinguish assignment, contact, engagement, and completion before defining the evaluation population and interpreting financial results.
Worked example: A trial assigns 500 members to an outreach offer, but only 200 engage. Comparing all 500 with the randomized control group estimates the offer's effect. Comparing the 200 engaged members alone may select people with systematically different health behaviors and access.
Common mistake: Removing nonengaged intervention members from a randomized comparison can create selection bias even though the original assignment was randomized.
Official syllabus reference: [2]
28. Palliative-care evaluation and immortal-time bias
Palliative-care evaluations can assess symptoms, quality of life, utilization, survival, and spending; lower spending alone is not a complete outcome. Defining recipients by a service received after cohort entry can create immortal-time bias if members must survive long enough to receive that service. Align eligibility, treatment assignment, and follow-up time, or use methods appropriate for time-varying exposure. Propensity adjustment addresses measured baseline differences but does not automatically correct a misaligned time origin. Prognosis, goals, and clinical severity can confound observational comparisons, so interpret estimated differences within the design's limits.
Worked example: A study labels anyone receiving palliative care within six months as exposed but starts follow-up at diagnosis. Recipients necessarily survive until their first visit. Assigning that earlier survival time to the exposed group can create an artificial survival advantage unrelated to care.
Common mistake: Assuming propensity matching removes immortal-time bias overlooks the separate need to align when eligibility, exposure assignment, and outcome observation begin.
Official syllabus reference: [2]
29. Pharmacy net cost and financing incentives
Pharmacy economics distinguishes list prices, negotiated ingredient payments, member cost sharing, rebates, fees, and final net payer cost. Financial arrangements can distribute these amounts differently across plans, intermediaries, pharmacies, and manufacturers. A high rebate does not necessarily imply the lowest net cost or best clinical value. Compare clinically appropriate alternatives using consistent net-cost definitions and the relevant contractual perspective. Rebate timing, guarantees, formulary incentives, utilization, and specialty-drug exposure can affect measured results. A financing change should be evaluated separately from any accompanying change in medication access, adherence, or health outcomes.
Worked example: Two clinically appropriate options cost the plan $500 and $420 before rebates. The first returns $100 and the second $10, giving net costs of $400 and $410. Fees, adherence differences, and member costs could still alter the overall value comparison.
Common mistake: Selecting the product with the largest rebate without comparing net cost and clinical consequences can increase spending or create conflicting incentives.
Official syllabus reference: [2]
30. Savings uncertainty and overlapping interventions
Estimated savings contain sampling variation and uncertainty from trend, selection, risk adjustment, claims completion, and model choice. Sensitivity analysis changes defensible assumptions to show how conclusions depend on them; interval estimates address statistical uncertainty under a specified model. These are complementary tools. When members receive several interventions, adding each program's estimated savings can double count shared avoided events. Define whether the objective is a program's marginal contribution or a combined portfolio effect. Attribution rules should reflect the evaluation design rather than allocating the same outcome independently to every participating vendor.
Worked example: A diabetes program and a transition-of-care program both claim savings from the same avoided readmission. A portfolio evaluation measures the combined difference once. Estimating either program's incremental contribution requires a comparison that distinguishes its presence from the other program's services.
Common mistake: Adding vendor-reported savings across overlapping populations can exceed the portfolio's actual counterfactual difference because the same avoided utilization is credited repeatedly.
Official syllabus reference: [2]
Healthcare Risk Adjustment
31. Risk normalization and common-population comparisons
Risk normalization adjusts experience to a common morbidity basis so that measured population differences do not dominate comparisons. Under a proportional model, normalized cost equals observed cost multiplied by target risk divided by observed risk. This relationship is an assumption, not a universal property of every risk score. Scores must be comparable in outcome, calibration, population, and period. Normalization does not automatically remove provider price, benefit, or access differences. Use it to support pricing, experience analysis, or performance comparison while identifying residual factors that could explain differences after morbidity adjustment.
Worked example: A population costs $720 PMPM at mean risk 1.20. Normalizing to risk 1.00 gives $600 PMPM under proportionality. Comparing that figure with another population requires compatible scores and consideration of different benefits, negotiated prices, and service coverage.
Common mistake: Multiplying observed cost by its own risk score instead of dividing to a common basis amplifies morbidity differences rather than normalizing them.
Official syllabus reference: [2]
32. Prospective versus concurrent risk models
A prospective risk model uses information from an earlier period to predict a later outcome, while a concurrent model uses information from the outcome period itself. Concurrent models can explain current spending well because diagnoses arise during costly care, but that information may be unavailable when prospective decisions are made. Match model timing to payment, pricing, or care-management use. Enrollment gaps and diagnosis lookback periods affect both approaches. Compare performance only with aligned prediction horizons and available information, and distinguish explanatory fit from genuine advance prediction of future resource needs.
Worked example: A plan setting next year's budget can use this year's diagnoses prospectively. A concurrent model incorporating next year's hospitalization diagnoses may explain next year's claims more closely, but those diagnoses were unavailable when the original budget decision had to be made.
Common mistake: Using superior concurrent fit to claim superior advance forecasting ignores that some predictors were observed only during the period being predicted.
Official syllabus reference: [2]
33. HCC hierarchies and interaction terms
Hierarchical condition category models map eligible diagnoses into categories and apply specified relationships among related conditions. A hierarchy can suppress a less severe category when a more severe one is present, avoiding designated double counting. Other conditions may contribute independently, and specified interactions can represent additional joint effects. Apply mappings, hierarchies, and interactions in the model's required sequence. A score is not the sum of every diagnosis coefficient. Actual categories and coefficients depend on model version, population, and payment context, so illustrative examples should not be substituted for operational specifications.
Worked example: An illustrative model has demographic contribution 0.60 and related disease contributions 0.40 and 0.20, retaining only the more severe disease. The subtotal is 1.00. An unrelated eligible condition could still add its own contribution if the model permits it.
Common mistake: Suppressing every additional diagnosis after identifying one severe condition misapplies hierarchy logic, which concerns specified relationships rather than all comorbidity.
Official syllabus reference: [2]
34. Model-version bridges and score comparability
Risk scores can change when diagnosis mappings, coefficients, hierarchies, calibration populations, or normalization conventions change. A raw score difference across versions therefore combines model effects with any genuine population change. Build a version bridge by scoring the same fixed population under both specifications, then separately examining enrollment and morbidity changes. Maintain effective dates and reproducible mappings. Payment implementation can include additional rules beyond the clinical model itself. A current syllabus containing historical technical papers does not make every historical model provision current for operational use; the applicable program year must be identified.
Worked example: A fixed cohort scores 1.08 under an older model and 1.02 under a newer model. That six-point change reflects model application to unchanged records. A separate comparison using current members is needed before attributing any remaining change to population morbidity.
Common mistake: Interpreting a lower score after a model update as improved health confuses recalibration with a change in members' clinical conditions.
Official syllabus reference: [2]
35. Medicaid rate cells and budget-neutral relativity
Medicaid risk adjustment often operates within defined eligibility or rate-cell structures, where population needs and benefit responsibilities can differ materially. Relative risk can redistribute a fixed payment pool among plans or cohorts, with a base rate selected to preserve the designated budget. Exposure weighting is essential. Do not pool incomparable eligibility groups simply because scores share a numerical scale. Distinguish adequacy of the total capitation budget from fairness of its allocation: budget-neutral risk adjustment changes distribution but does not by itself correct an insufficient aggregate rate. State-specific program specifications determine actual implementation.
Worked example: A fixed monthly pool of $1.2 million covers 1,000 members with total risk-weighted exposure of 1,100. A base of approximately $1,090.91 per risk unit preserves the budget. Whether $1.2 million is adequate requires a separate assessment of expected covered costs.
Common mistake: Assuming budget neutrality establishes adequate Medicaid capitation confuses redistributing available funds with determining the total resources needed for covered care.
Official syllabus reference: [2]
36. ACA transfers and relative market risk
ACA risk adjustment seeks to reduce incentives associated with enrolling healthier rather than sicker members within specified markets. Transfers depend on relative modeled risk and additional program factors, rather than simply reimbursing each plan's actual claims. A plan can have high costs because of prices or administration without having correspondingly high compensable morbidity. Member exposure and the applicable transfer pool matter when reconciling receipts and payments. Simplified budget-neutral examples illuminate redistribution but omit operational formula details. Use the applicable benefit-year methodology for real calculations rather than extrapolating from a historical technical paper.
Worked example: Two plans have equal enrollment, but one has higher modeled morbidity. A simplified pool transfers funds toward that plan while balancing the other plan's payment. If both scores rise identically, relative position may remain unchanged even though their raw scores increased.
Common mistake: Treating risk adjustment as reimbursement of every plan's excess claims overlooks its focus on modeled relative risk and defined transfer factors.
Official syllabus reference: [2]
37. Benefit richness, induced utilization, and morbidity
Observed spending reflects both health status and insurance design. Richer benefits can increase the insured share of spending and may change service use through lower member prices, even with unchanged underlying morbidity. Risk-adjustment analysis must distinguish these effects from health differences and apply the relevant program's treatment of benefit design. A score calibrated to total spending is not automatically a prediction of plan liability under every benefit. In pricing, separate morbidity, utilization responses, covered services, and actuarial value rather than using one adjustment factor to represent all of them.
Worked example: Two otherwise similar populations have different deductibles. The lower-deductible group may use more outpatient services and generate a larger insurer-paid share. Those differences do not necessarily indicate worse health, so a comparison based only on paid claims can misstate relative morbidity.
Common mistake: Interpreting higher insurer-paid claims under richer coverage as entirely greater illness mixes benefit mechanics with the health-status differences risk adjustment aims to measure.
Official syllabus reference: [2]
38. High-cost pooling and market stabilization
Risk adjustment, reinsurance, and other stabilization mechanisms address related but distinct financial problems. Risk adjustment reallocates revenue for measured risk differences; reinsurance reimburses specified portions of large claims; aggregate sharing mechanisms can address broader financial deviations. Extremely costly cases may remain poorly predicted even in a useful morbidity model. Combining mechanisms requires clear sequencing and consistent treatment of recoveries to avoid duplicate compensation. Evaluate incentives and residual volatility separately from average predictive fit. Historical stabilization arrangements should not be assumed to remain available in every current market or benefit year.
Worked example: A plan receives a morbidity-based transfer and a separate recovery for a qualifying large claim under a hypothetical reinsurance arrangement. The analyst must follow the specified treatment of that recovery when measuring costs, rather than assuming the two mechanisms independently cover the same loss.
Common mistake: Calling every high-cost claim recovery risk adjustment obscures the different triggers, financing sources, and incentives of risk scoring and reinsurance.
Official syllabus reference: [2]
39. Coding intensity versus true morbidity
Diagnosis-based risk scores depend on documented and accepted clinical information, so coding practices can change measured risk without equivalent changes in health. More complete capture may correct omissions, while unsupported diagnoses can overstate payment risk. Analyze score changes using enrollment, disease incidence, encounter frequency, documentation practices, and model-version bridges. Apply the relevant program's eligibility and submission rules rather than assuming any diagnosis is usable. Distinguishing coding intensity from morbidity matters for pricing, normalization, and provider incentives. A score increase alone cannot establish that medical need or expected cost increased by the same percentage.
Worked example: A plan's score rises after a documentation initiative, while clinical outcomes and enrollment remain stable. The analyst separates newly captured established conditions from incident disease and checks support for submitted diagnoses before projecting an equal increase in expected medical claims.
Common mistake: Assuming every newly recorded diagnosis represents newly developed illness can overstate trend and misattribute changes in payment to worsening health.
Official syllabus reference: [2]
40. ASOP 23 and fitness of risk-adjustment data
Data-quality work asks whether available information is suitable for the intended actuarial use, rather than assuming that a large dataset is accurate. The syllabus assigns ASOP 23, making data selection, review, limitations, and reliance relevant to risk adjustment. Check enrollment exposure, diagnosis completeness, duplicate records, impossible values, coding changes, and reconciliations to source totals. Investigate anomalies in proportion to their potential effect. Data review is not automatically an audit. Material unresolved limitations can affect the method, interpretation, or communication even when correcting every individual record is impractical.
Worked example: One plan's diagnosis count falls sharply after a file-layout change while enrollment stays stable. Before interpreting lower risk as improved morbidity, the actuary reconciles rejected records and source totals, quantifies the potential score impact, and communicates any unresolved completeness limitation.
Common mistake: Describing a dataset as audited merely because reasonableness checks were performed overstates the assurance provided and can mislead users of the analysis.
Official syllabus reference: [2]
41. ASOP 45 and model suitability for intended use
A health-status risk-adjustment methodology should be evaluated for the population, outcome, and decision it is intended to support. The syllabus assigns ASOP 45 in this context. Relevant considerations include calibration data, available predictors, prediction timing, benefit scope, model performance, and implementation limitations. A model useful for payment allocation may not be equally useful for identifying members who benefit from outreach. Validate aggregate and subgroup results, not just one summary statistic. The actual standard provides the governing professional guidance; the syllabus outline establishes relevance without reproducing its detailed requirements.
Worked example: A commercial model calibrated on working-age adults is proposed for a Medicaid population with different eligibility groups. Before use, the actuary examines population fit, subgroup calibration, available diagnoses, benefit differences, and whether recalibration or a different methodology better serves the payment objective.
Common mistake: Assuming a commercially licensed model is suitable for every population substitutes vendor availability for evaluation of its intended use and limitations.
Official syllabus reference: [2]
42. ASOP 41 and reproducible actuarial communication
An actuarial communication should enable intended users to understand the result, its purpose, and material limitations. In risk adjustment, essential context includes population, exposure period, model version, prediction horizon, score normalization, data sources, and significant assumptions. Distinguish observed experience from modeled estimates and identify reliance where relevant. The syllabus assigns ASOP 41, but its exact provisions require the standard itself. Communication should support appropriate use without overwhelming the audience with irrelevant detail. A technically correct number can still mislead when its denominator, version, or decision boundary is left unstated.
Worked example: A report states that average risk is 1.12 but also identifies the model year, eligible member-months, diagnosis period, and reference population. It explains that the score supports a payment comparison and should not be read as an individual probability of hospitalization.
Common mistake: Reporting a risk score without its reference basis encourages users to compare incompatible models or mistake a relative index for an absolute probability.
Official syllabus reference: [2]
Medical Data
43. Clinical terminology, incidence, and prevalence
Clinical interpretation begins by distinguishing a disease, a symptom, a diagnostic test, and a treatment. Claims may record these at different stages of an episode, so a test does not necessarily confirm the disease being investigated. Incidence measures newly occurring cases over an appropriate population and time; prevalence measures existing cases in a defined population at a point or period. Claims-based identification depends on coding and observation history rather than direct biological onset. Disease duration, survival, enrollment, and detection can change prevalence even when incidence is stable, affecting actuarial projections.
Worked example: A plan observes 100 members with diabetes among 2,000 continuously enrolled adults, a 5% observed prevalence. Twenty newly qualifying cases during the year do not establish biological incidence unless prior disease history and the case-definition lookback are sufficient to distinguish new from previously unobserved disease.
Common mistake: Treating the first diagnosis in an insurer's file as the patient's disease onset ignores earlier history outside the available observation window.
Official syllabus reference: [2]
44. ICD-10-CM diagnoses versus ICD-10-PCS procedures
ICD-10-CM represents diagnoses in U.S. clinical coding, while ICD-10-PCS represents procedures in the U.S. inpatient hospital setting. They answer different questions and should not be treated as interchangeable because both contain ICD-10 in their names. Diagnosis interpretation can depend on specificity, encounter context, and additional coding instructions. Procedure analysis also requires setting and version awareness. A diagnosis code alone does not establish severity, treatment, or payment eligibility beyond the relevant rules. Maintain the original code and effective date when building longitudinal disease definitions or reimbursement analyses.
Worked example: An inpatient record includes a diagnosis describing a condition and a separate procedure code describing what was done. A utilization study seeking operations should use the applicable procedure information rather than count every diagnosis of the condition as evidence of surgery.
Common mistake: Searching only diagnosis fields to count completed procedures confuses the reason for care with the intervention actually performed during the encounter.
Official syllabus reference: [2]
45. CPT, HCPCS, modifiers, and service units
CPT identifies many professional and outpatient services, while HCPCS Level II includes additional supplies, drugs, equipment, and other services. Claim interpretation often requires modifiers and units as well as the base code. A unit may represent a visit, time interval, dose quantity, or another specified increment, so counts are not automatically comparable across codes. Payment rules may bundle components or alter allowances for circumstances identified by modifiers. Preserve setting, date, and coding version, and distinguish submitted service units from distinct encounters or members when calculating utilization and average prices.
Worked example: A drug claim reports ten units where each unit represents a specified dose increment. Counting that line as ten treatment visits would misstate utilization. The analyst uses the code's unit definition and encounter information to distinguish administered quantity from the number of visits.
Common mistake: Summing units across unrelated procedure codes as though every unit represents one visit creates a utilization measure without a consistent clinical meaning.
Official syllabus reference: [2]
46. Revenue codes versus DRG case classifications
Revenue codes identify facility service categories or departments on institutional claim lines, while a DRG groups an inpatient case under a specified classification system. A revenue code is therefore line-oriented and a DRG case-oriented. Neither necessarily identifies every clinical detail needed for analysis. Combining revenue, procedure, diagnosis, and admission information can distinguish settings and services more accurately. Avoid summing a case-level payment repeatedly across its component lines. Reimbursement interpretation requires the actual payment method, because a facility may report detailed charges even when the contract pays a bundled case amount.
Worked example: An inpatient claim contains room, laboratory, and pharmacy revenue lines but receives one DRG-based payment. The analyst uses the revenue lines to describe service composition and the DRG to classify the admission, without treating each line as a separately paid inpatient case.
Common mistake: Multiplying a case-level DRG payment by the number of facility claim lines duplicates reimbursement and inflates estimated hospital spending.
Official syllabus reference: [2]
47. Code-version crosswalks and longitudinal comparability
Coding systems change through additions, deletions, revised descriptions, and changed classification rules. A crosswalk can relate versions, but mappings may be one-to-many or depend on clinical detail unavailable in older data. Apparent disease or procedure trends can therefore reflect classification changes rather than real utilization. Retain service dates and code versions, document mappings, and test breaks around implementation dates. Use broader stable groupings when justified, while acknowledging lost specificity. A forced one-to-one mapping can create false precision and distort both longitudinal prevalence estimates and reimbursement projections.
Worked example: One broad historical code splits into three specific codes. The analyst groups the three new codes with the old category to measure a continuous trend. Estimating each new subtype historically would require information not contained in the original broad code.
Common mistake: Assigning every old broad code to a single new subtype invents clinical detail and can create an artificial shift in disease mix.
Official syllabus reference: [2]
48. Common-rule repricing and coding-driven reimbursement
Common-rule repricing applies one payment schedule and classification basis to different periods or populations. It helps separate contractual price changes from differences in service volume, mix, and coding. When grouping rules also change, rerunning comparable records through a common grouper can identify classification effects, subject to available detail. Repricing does not eliminate genuine clinical differences or repair incomplete source records. Reconcile total allowed amounts, adjustments, and outlier treatment before interpreting the bridge. Keep actual historical payments distinct from standardized analytic values, which answer a hypothetical comparison question.
Worked example: Historical claims paid $4 million and current claims paid $4.6 million. Repricing both at current fees yields $4.4 million and $4.6 million. This suggests part of the observed increase reflects price changes, while the remaining difference still requires utilization and coding analysis.
Common mistake: Calling the entire actual spending increase a utilization trend overlooks fee changes and classification effects that common-rule repricing is designed to isolate.
Official syllabus reference: [2]
49. Clinical policy bulletins and coverage criteria
A clinical policy bulletin describes a payer's assessment or criteria for particular services, often addressing medical necessity, indications, evidence, or exclusions. Interpretation requires the effective version, eligible population, service definition, and interaction with the member's benefit coverage. A listed procedure code does not by itself establish that every use is covered. Separate clinical appropriateness, benefit inclusion, authorization, and payment rules. For actuarial analysis, translate applicable criteria into measurable eligibility assumptions while identifying clinical details missing from claims. Avoid treating one payer's bulletin as a universal clinical or legal standard.
Worked example: A bulletin covers a procedure for a specified indication after documented prior treatment. Counting every claim with that procedure code as policy-eligible would ignore the indication and treatment-history requirements. A coverage estimate may need chart information or explicit uncertainty assumptions.
Common mistake: Assuming a code listed in a policy bulletin guarantees reimbursement overlooks indication-specific criteria, benefit terms, and other applicable claim requirements.
Official syllabus reference: [2]
50. Public, private, and linked healthcare data
Public data can provide population benchmarks, geography, or aggregated utilization, while private claims, enrollment, and clinical records often provide more detailed member-level experience. Availability does not determine fitness for purpose. Compare populations, coverage scope, measurement periods, suppression, sampling, and definitions before combining sources. Linking records requires reliable identifiers or evaluated matching methods, and unmatched records can create selection bias. Claims measure billed or paid services rather than every clinical event, while clinical systems may omit care outside their network. Preserve privacy and permitted-use boundaries when choosing linkage and reporting detail.
Worked example: A plan combines its claims with a public county deprivation measure and a hospital registry. County data add context, while the registry adds severity. The analyst checks address dates, registry coverage, and unmatched members before interpreting any association with spending.
Common mistake: Assuming linked data are automatically more representative ignores the possibility that only well-documented, stable, or network-connected members successfully match.
Official syllabus reference: [2]
51. Study designs and triangulation of evidence
Randomized trials compare assigned interventions, cohort studies follow exposed groups, case-control studies sample by outcome, and cross-sectional studies assess variables within a defined snapshot. These designs support different inferences and face different biases. Identify the target question, population, exposure timing, outcome, and comparison before ranking evidence. Randomization strengthens causal identification but does not ensure applicability to every local population. Triangulation uses sources with different strengths and weaknesses to test a coherent explanation. Agreement is informative only after aligning definitions; averaging incompatible estimates can conceal rather than resolve their differences.
Worked example: A trial estimates coaching effects among volunteers, claims show local admission patterns, and a registry describes disease severity. Together they inform transferability and evaluation design. Their estimates should not simply be averaged because they measure different populations and may use different outcome definitions.
Common mistake: Treating every retrospective database analysis as a case-control study confuses data collection timing with the sampling design that defines the study.
Official syllabus reference: [2]
Social Determinants of Health
52. Social pathways from living conditions to spending
Social determinants influence health through material resources, environmental exposure, stress, access, and opportunities for prevention or treatment. Housing instability, transportation barriers, and food insecurity can affect different points in a care pathway rather than acting as interchangeable risk markers. Spending effects can run in either direction: unmet need may suppress current utilization while increasing later complications. Specify the proposed mechanism before selecting an intervention or interpreting a model coefficient. Association with cost does not establish that changing the measured social variable would produce the same effect, especially when structural confounding remains.
Worked example: A member misses routine appointments because transportation is unreliable. Providing rides may increase outpatient spending initially while improving access and potentially preventing later acute events. An evaluation should measure completed visits and health outcomes rather than presume immediate medical savings.
Common mistake: Assuming social support must immediately reduce claims overlooks unmet need that can produce appropriate increases in preventive and outpatient service use.
Official syllabus reference: [2]
53. Area deprivation and ecological inference
Area-level social measures describe a geographic population and can capture neighborhood context, but they do not directly establish an individual's circumstances. Geographic aggregation, boundary choice, and address quality affect the measure. Associations can reflect both contextual influences and the characteristics of residents. Use explicit labels distinguishing area attributes from individual observations, and align geographic vintage with the analysis period. Area measures can support resource allocation or predictive modeling when validated, but individual service eligibility may require more direct information. Avoid interpreting an area association as a personal diagnosis or causal effect.
Worked example: A member lives in a tract with low average vehicle access. That information can support community transport planning, but the member may own a car. Outreach can offer a needs assessment without recording lack of transportation as an established individual fact.
Common mistake: Assigning every resident the tract's average social condition creates false individual information and can misdirect assistance or distort model interpretation.
Official syllabus reference: [2]
54. Housing interventions and cost-consequence analysis
Housing interventions can produce housing stability, safety, health, and financial outcomes that do not fit a single medical-savings measure. Cost-consequence analysis reports relevant costs and outcomes separately, while ROI focuses on monetized benefits from a defined perspective. Include placement, support, administration, and implementation costs, and use a credible counterfactual for medical spending. Distinguish enrolled, offered, housed, and continuously housed populations. Conditioning only on successful placement can introduce selection. A payer's negative medical ROI can coexist with important benefits to members or other public systems, without proving the intervention is cost-effective.
Worked example: A housing program costs $900,000 and estimates $700,000 in avoided medical spending. It also reports more stable housing days. The payer's net medical benefit is negative $200,000, while the housing result remains a separate outcome requiring explicit valuation or cost-consequence reporting.
Common mistake: Declaring a housing program ineffective solely because medical savings do not exceed costs ignores outcomes outside the payer's narrow financial perspective.
Official syllabus reference: [2]
55. Social-risk adjustment and payment incentives
Including social-risk variables in payment models can recognize resource needs not fully represented by diagnoses. The policy objective must be explicit: supporting access, improving allocation, or reducing incentives to avoid disadvantaged populations. Predicting observed spending alone can reproduce underuse where unmet need suppresses claims. Payment adjustment and quality assessment also require different judgments; an adjustment intended to fund additional care should not automatically lower expectations for equitable outcomes. The syllabus assigns MassHealth model readings, but their detailed variables and coefficients require those documents rather than inference from their titles.
Worked example: A plan serving members with housing instability has low routine utilization but frequent emergency care. A social-risk payment analysis examines whether additional resources support needed services, rather than assuming low observed outpatient spending demonstrates low need or warrants reduced funding.
Common mistake: Using social adjustment to normalize poor outcomes without examining inequity can conceal barriers that the additional payment was intended to help address.
Official syllabus reference: [2]
56. Informative missingness in social-needs screening
Missing social-needs data often reflect who was contacted, who trusted the process, language access, service setting, or willingness to disclose. Missingness is therefore not necessarily random or equivalent to absence of need. Compare screening availability and response across groups and care settings. Preserve distinctions among not asked, declined, unknown, and reported no need where data permit. Imputation requires assumptions and can obscure systematic nonresponse. For service delivery, collect only information needed for a legitimate purpose, explain its use, and avoid making access depend on unnecessary disclosure of sensitive circumstances.
Worked example: A plan screens members during clinic visits, so people with transport barriers respond less often. Coding all unscreened members as having no transport need reverses the likely pattern. The analyst retains missingness categories and evaluates outreach methods that reach less-connected members.
Common mistake: Replacing every missing social-needs response with zero assumes absence of need and can systematically understate risk among members least likely to be screened.
Official syllabus reference: [2]
57. Incremental validity and fairness of SDOH predictors
Evaluate an SDOH predictor by whether it improves the intended decision beyond existing information, not merely whether its coefficient is statistically significant. Compare calibration, error, and operational outcomes with and without the variable using appropriate validation data. Examine subgroup performance and the consequences of underprediction or overprediction. Geographic proxies can encode structural disadvantage, and excluding explicit social variables does not remove information embedded in other predictors. Distinguish predictive usefulness from causal interpretation. A model that predicts spending accurately may still allocate insufficient care to populations whose historically low spending reflects restricted access.
Worked example: Adding neighborhood deprivation improves aggregate cost prediction but leaves underprediction among members with limited access. The actuary checks subgroup calibration and tests whether the outreach allocation changes appropriately, rather than accepting a small overall accuracy gain as proof of equitable usefulness.
Common mistake: Assuming removal of an explicit social variable eliminates bias overlooks correlated geographic, utilization, and demographic features that can preserve the same patterns.
Official syllabus reference: [2]
Predictive Analytics
58. Choosing cost, event-risk, and benefit models
A predictive model's target should match the business decision. Expected-cost models estimate spending, event models estimate outcomes such as admission, and treatment-benefit models seek differences under alternative interventions. High expected cost does not imply high preventable risk or large intervention benefit. Models can also be prospective or concurrent, and diagnosis-based or based on other available information. Define the outcome, horizon, population, and action before choosing predictors or performance metrics. A relative payment score is not automatically a calibrated individual probability, and a ranking score may require additional calibration before probability-based decisions.
Worked example: One member has high expected spending from a necessary treatment with little modifiable utilization. Another has lower total expected cost but a preventable readmission risk responsive to outreach. A cost-ranking model and an intervention-benefit model can reasonably prioritize different members.
Common mistake: Using the highest cost score as proof of greatest care-management benefit confuses prediction of spending with prediction of an intervention's incremental effect.
Official syllabus reference: [2]
59. Claims runout, timestamps, and temporal leakage
Healthcare data have several relevant times: when care occurred, when information was documented, when a claim arrived, and when it became available for modeling. A feature can predate an outcome clinically yet still be unavailable at the operational prediction time. Claims runout also makes recent outcomes incomplete. Reproduce the deployment information set during development, separate observation and prediction windows, and allow adequate outcome completion. Validate on later periods to test temporal transportability. Keep repeated member records from creating inappropriate information sharing across splits, and investigate coding or system changes that alter feature availability.
Worked example: A January discharge model uses December services, but some December claims do not arrive until March. A retrospective extract contains them and can inflate performance. A realistic January prediction must use only records actually available then or an explicitly supported data-lag approach.
Common mistake: Filtering predictors only by service date overlooks reporting delays and can leak information that the production system would not yet possess.
Official syllabus reference: [2]
60. Discrimination, calibration, capacity, and ethical use
Discrimination measures ranking, while calibration compares predicted and observed outcomes. Neither alone establishes that a model improves a healthcare decision. Evaluate positive predictive value, recall, expected utility, or other measures at the actual intervention capacity and consider the costs of false positives and false negatives. Validate subgroup performance and whether the target, such as historical spending, is an appropriate proxy for need. Model deployment also requires a feasible action and evaluation of its effects. A model can remain statistically accurate while becoming operationally harmful if access, population, or intervention effectiveness changes.
Worked example: An outreach team can contact 200 members monthly. Two models have similar AUROC, but one identifies more eventual admissions among its top 200 and has better subgroup calibration. Those differences matter directly to workload allocation, although outreach benefit still requires separate evaluation.
Common mistake: Selecting a model solely by AUROC ignores probability accuracy, real staffing limits, unequal error consequences, and whether the resulting intervention benefits members.
Official syllabus reference: [2]
Official syllabus sources
- GH 301 Health Analytics and Management
- Syllabus with Learning Objectives/Outcomes and Readings – Nov 2026/Mar 2027/Jul 2027