Risk adjustment — Expected vs actual cost

Anchor Health Cooperative

Risk adjustment

A synthetic demo surface — "Anchor Health Cooperative" is a fictional, obviously-fake HMO. No real payer, no PHI. Every number here reconciles to the vault: the risk score is cs_member_risk_score (clinical diagnoses → chronic risk categories); actual paid = Σ fct_claim (status = paid) in 2017–2026. The app carries no business logic — the expected-cost model and residual are computed in the fct_expected_cost gold mart.

Expected vs actual cost · the over/under-performer signal

The payer's central "how expensive should this member be?" question, answered from clinical risk. Each member's risk score drives an expected cost — floor + marginal × risk_score, a two-parameter fit whose constants are recomputed at build from the population (the floor is the mean paid of zero-risk members, so a member with real utilization but no chronic diagnoses still expects a non-zero cost). The residual = actual − expected is the signal: who costs more (or less) than their risk says.

Member-resolved members

675

Total expected (model)

$1,090,657

Total actual (paid claims)

$1,090,657

Expected ≈ actual by construction. The marginal is solved so the population's total expected equals total actual — the model reallocates cost across members, it never invents or destroys any. Total expected ÷ total actual = 1.000 (≈ 1.000). The residual is where the story lives, not the total.

Cost rises with risk — but the fit is honestly noisy

Average actual paid climbs steadily across the risk-score bands: the score is genuinely predictive. It is not precise — the residual within each band is large (R²≈0.18), and that spread is the feature, not a bug. It is exactly the "who costs more than their risk says" signal a payer chases.

No Results

The residual scatter — expected vs actual, per member

Each dot is one member: risk-driven expected cost on the x-axis, actual paid on the y-axis. Members on the diagonal cost what their risk predicts; dots high above it are under-performers (cost more than expected — the cost-overrun tail); dots below are over-performers (cost less).

Top under-performers — cost more than risk predicts

The cost-overrun tail: members whose actual paid claims most exceed their risk-driven expectation (largest positive residual). This is the list a care-management program would triage.

No Results

Top over-performers — cost less than risk predicts

The mirror image: members whose actual paid claims fall furthest below expectation (largest negative residual) — high apparent risk, low realized cost.

No Results

Reconciles to the vault (check it live): risk score = the ratified cs_member_risk_score (chronic diagnoses → risk categories), bridged to the payer member via cs_patient_member_same_as; actual paid = Σ fct_claim.total_paid_amt where claim_status = 'paid' in 2017–2026 (the same population + window as the loss-ratio mart). The floor and marginal are recomputed at build — no hardcoded constants.