SEER-Medicare Linked Data

SEER-Medicare

links the NCI SEER cancer registries — population-based tumor registries capturing all incident cancers in defined geographic areas with standardized stage (SEER historic, TNM-derived), morphology, and first-course treatment abstracted from medical records — to Medicare FFS claims (Parts A/B, optionally Part D since 2006) for beneficiaries residing in SEER catchment areas. The linkage is performed by NCI; researchers receive de-identified linked files.

Why it matters for RWE

Registry data supplies what claims cannot: validated cancer diagnosis (histology confirmed), stage at diagnosis, and chart-abstracted first-course treatment. Claims supply what registries cannot: longitudinal utilization, recurrence-era services, comorbidity burden, Medicare payments, and death dates. The combination supports treatment-pattern, cost, comparative-effectiveness, and survivorship research in a population-based elderly cohort — the backbone of US cancer health-services RWE.

Operational characteristics

• Population: patients 65+ with FFS coverage (registry captures all ages but claims exist only for Medicare enrollees); SEER areas cover roughly 30-47% of the US population depending on era.
• Registry variables: month/year of diagnosis (not day), SEER stage, grade, morphology codes, first-course surgery/radiation/chemotherapy flags.
• Claims variables: ICD-9/10 diagnoses, CPT/HCPCS procedures, revenue codes, Part D drug events, payment amounts, date of death.
• FFS requirement: MA enrollees lack claim detail; analyses must restrict to FFS or handle MA months explicitly.

Common pitfalls

• Day-of-diagnosis missing: use month-level windows; avoid misclassifying pre-diagnosis services.
• Lookback limited to claims history: Medicare eligibility begins ~65; comorbidity lookback before 65 is impossible for most.
• Recurrence invisible: no registry flag for recurrence; algorithms using claims after a disease-free window are approximations needing validation.
• Generalizability: SEER areas are not nationally representative; MA growth over time shrinks the analyzable FFS denominator.

Pros, cons, and trade-offs

• vs pure claims (MarketScan/Medicare): cancer validation, staging, and cause-specific context vs broader populations without clinical depth.
• vs Flatiron-style clinico-genomic EHR: population-based incidence capture vs deep genomic/clinical detail in treated cohorts.
• Trade-off: elderly-only claims coverage vs pediatric/adult registry breadth.

When NOT to use

Under-65 oncology RWE (no claims); questions requiring recurrence adjudication beyond validated algorithms; biomarker-driven treatment effects before genomics-linked expansions.

Decision diagram

flowchart LR
  R[SEER registries\ntumor - stage - first course] --> LK[NCI linkage]
  M[Medicare FFS claims A/B/D] --> LK
  LK --> SMD[De-identified SEER-Medicare files]
  SMD --> P[Treatment patterns - costs - outcomes RWE]

Worked example

Scenario

Compare 5-year total Medicare costs after diagnosis between stage III colon cancer patients receiving adjuvant chemotherapy versus not.

Dataset

Cost comparison adjusted for demographics and comorbidity.

Steps

Result

Adjuvant recipients incurred 18% higher adjusted 5-year costs, driven by first-year treatment-period spending; sensitivity to MA-month exclusion was minimal.

Trade-offs

Runnable example

\
import pandas as pd

def adjuvant_flag(carrier, pde, resection_date, window=120):
    # Part B J-codes + Part D chemo within window days post-resection
    b = carrier[(carrier.jcode.isin(CHEMO_JCODES)) &
                (carrier.svc_date >= resection_date) &
                (carrier.svc_date <= resection_date + pd.Timedelta(days=window))]
    d = pde[(pde.ther_class=="antineoplastic") &
            (pde.fill_date >= resection_date) &
            (pde.fill_date <= resection_date + pd.Timedelta(days=window))]
    return int(len(b) > 0 or len(d) > 0)

def cumulative_costs(inpatient, carrier, pde, dx_date, months=60):
    end = dx_date + pd.DateOffset(months=months)
    tot = {f: f[(f.date >= dx_date) & (f.date < end)].payment.sum()
           for f, name in [(inpatient,"ip"),(carrier,"op"),(pde,"pd")]}
    return sum(tot.values())

\
library(data.table)

CHEMO_J <- c("J9206","J9263","J8520","J8521")

adjuvant_flag <- function(carrier, pde, resection_dt, window = 120L) {
  b <- carrier[JCODE %in% CHEMO_J &
               svc_date >= resection_dt &
               svc_date <= resection_dt + window, .N]
  d <- pde[ther_class == "antineoplastic" &
           fill_date >= resection_dt & fill_date <= resection_dt + window, .N]
  as.integer(b + d > 0)
}

cumulative_costs <- function(claims_list, dx_dt, months = 60L) {
  end <- as.Date(dx_dt) %m+% months(months)
  sapply(claims_list, function(f)
    f[date >= dx_dt & date < end, sum(payment)], USE.NAMES = FALSE) |> sum()
}

\
/* Adjuvant chemotherapy flag: any Part B chemo J-code or Part D antineoplastic */
proc sql;
  create table adj_b as
    select distinct patient_id, 1 as adj_flag
    from carrier
    where jcode in ('J9206','J9263','J8520','J8521')
      and svc_date between resection_date and resection_date + 120;

  create table costs60 as
    select patient_id,
           sum(case when date >= dx_date and date < intnx('month',dx_date,60)
                    then payment else 0 end) as cost_5yr
    from all_claims
    group by patient_id;
quit;

Citations

1. [1]Warren JL, et al. Overview of the SEER-Medicare Data: Content, Research Applications, and Generalizability to the United States Elderly Population. Medical Care. 2002.
2. [2]Enewold L, et al. Updated Overview of the SEER-Medicare Data: Enhanced Content and Applications. JNCI Monographs. 2020.
3. [3]Snow TS, et al. Comparison of Population Characteristics in Real-World Clinical Oncology Databases in the US. medRxiv. 2023.

1. [4]National Cancer Institute. Healthcare Delivery Research Program: SEER-Medicare Linked Data Resource.