Insights — Cost Variation Across Providers & Geography
Methodology: descriptive adaptations of the Agrawal & Choudhary (KDD-DMH 2013) framework on separate FY2023 inpatient and CY2023 outpatient tracks. The outputs are exploratory and not causal.
The 3-tier coefficient-of-variation framework — CV_b, CV_p, CV_nb — is attributed to Agrawal & Choudhary, An Analysis of Variation in Hospital Billing Using Medicare Data (KDD-DMH 2013). Their state-normalized-billing notation and billing-payment association analysis are methodological context. Current outputs use different periods, service universes, payment fields, and aggregation rules, so they are adaptations rather than controlled historical reproductions.
Q1: Which services show the highest observed charge/payment variation?
Per service code (DRG inpatient / APC outpatient), for services with at least 30 valid providers, we compute three descriptive dispersion metrics:
CV_b— coefficient of variation of billed charges (what the provider asks for)CV_p— coefficient of variation of the recorded payment denominator: average Medicare payment inpatient, average allowed amount outpatientCV_nb— coefficient of variation of payment-normalized billing, using the service-level payment ratio; this is a descriptive normalization, not a residual or causal measure
Five rankings (Agrawal's framework)
This isn't one CV ranking — it's five descriptive views, one per axis adapted from the prior framework. The generated tables use the current FY2023/CY2023 service universes and should not be read as a controlled reproduction of historical tables:
Inpatient — Top 5 per ranking
Outpatient — Top 5 per ranking
CV_b vs CV_p — billing variation vs payment variation, side by side
Each point is one service. The plot compares billing variation (CV_b, x-axis) with payment variation (CV_p, y-axis) under the current denominators. The FY2023 inpatient maximum is 1.5306 among 315 eligible DRGs; the CY2023 outpatient maximum is 1.1156 among 64 eligible APCs. Agrawal's historical maximum is not a like-for-like benchmark because the service universe and eligibility design differ, so no decade trend is inferred.
Q2: Are there regional patterns in observed charge/payment per service?
For each selected top-ranked service we compute descriptive state-level B/P/NB indices on a log2 scale, using provider means aggregated within state-service cells. The controls show the union of the top services from five ranking views, not every observed service:
B_s = log2(μ_b,s / μ_b,us)— state billing vs US meanP_s = log2(μ_p,s / μ_p,us)— state payment vs US meanNB_s = log2[(μ_b,s / μ_b,us) / (μ_p,s / μ_p,us)]— a normalized billing index for visualization; color is relative to the service's US mean and does not establish overcharging or cost causality
Primary inpatient geographic outputs exclude Maryland because of its all-payer rate-setting context. The included-versus-excluded comparison is recorded in data/maryland_sensitivity.json. Pick a selected service to view its choropleth (B-map, P-map, NB-map side by side):
Inpatient DRG
Outpatient APC
Billing–payment association at the state-service level
The current correlation artifact is explicitly a state-level ecological Pearson correlation: provider means are aggregated into state-service cells, cells with fewer than five providers are omitted, and services need at least four eligible state cells. Each service row also records a 95% Fisher-z interval; small state counts remain unstable. Primary geographic summaries exclude Maryland. Current medians are 0.4432 inpatient across 241 services and 0.1088 outpatient across 58 services; maxima are 0.9793 and 0.4458. These values are not directly comparable with a provider-level historical statistic.
Inpatient
Median billing-payment corr: —
Range: — · Negative-correlation services: —
FY2023 state-ecological summary; historical aggregation differs
Outpatient
Median billing-payment corr: —
Range: — · Negative-correlation services: —
CY2023 state-ecological summary; no like-for-like historical baseline claimed
The scatter below plots each service's state-ecological correlation against its normalized-billing index. Negative and positive values describe this aggregation; they do not identify hospital pricing behavior, cost-of-care mechanisms, or causality.
5 lowest & 5 highest state-ecological correlations (inpatient)
Lowest
Highest
Cross-dataset provider proxy comparison
We join on provider_ccn and identify providers in the top quartile of provider-level mean charge-to-payment proxy in both settings. The inpatient denominator is average Medicare payment and the outpatient denominator is average allowed amount; the aggregation is an unweighted mean of valid row-level ratios with no service standardization. The resulting cohort is descriptive and non-causal. No comparable provider-level join was found in the sources reviewed; that is not an absolute novelty claim.
The scatter plots each dual-setting provider's inpatient proxy (x) against outpatient proxy (y). The highlighted upper-right group is a selected proxy cohort; service mix, eligible breadth, suppression, and the different denominators affect placement. It is not a driver, systemic behavior, or markup finding.
Top 10 providers by proxy values in both settings
← approach · next: Outliers →