OneCare Vermont: Data & Code

The generated result files and the public inputs behind the accounting. Everything the figures consume is here; the heavy raw archives are reproducible from the source links.

Plain CSV and JSON · regenerate the figures with build_figures.py in the parent directory (Python, matplotlib, pandas); the estimates themselves rebuild from analysis/ given the companion checkout — see the methodology note for which inputs are redistributed here and which must be downloaded. The shipped SVGs are transparent so they sit on the page's own background; set ONECARE_PREVIEW=<dir> to render opaque PNGs on black instead, which is the only way to proof figure text — matplotlib saves it as glyph outlines, not as characters.

Generated results

File	Contents
results_manifest.json	The authority for every published number. Current intercept-shifted fits for all three outcomes: actual, synthetic and gap series by year, donor weights, both placebo rankings, leave-one-pre-year-out diagnostics, donor sensitivity, predictor-set stability, the family-wise permutation across the three outcomes, and the stated limitations. The figures and the prose read from this file.
ma_robustness.json	Medicare Advantage robustness on the published specification: yearly-MA predictors and the ±10-point restricted donor pool, each with its own recomputed placebo rank, against the baseline.
scm_placebo_statistics.csv	Pre-period RMSPE, post-period RMSPE and their ratio for each of the 49 placebo states, per outcome, all from that same specification — the permutation distributions behind both rank columns, and the joint pre/post distribution plotted in figure 2b.
synth_results.json	Per-outcome series in a flatter shape than the manifest. The manifest is what the figures and prose read; prefer it.
vermont_gap_spending.csv	Vermont's Medicare spending gap (actual − synthetic) per beneficiary-year.
placebo_gaps_spending.csv	The same spending gap computed for all 49 placebo states — the permutation distribution behind p = 0.32.
hospital_margins.csv	Aggregate hospital total margins, Vermont / New Hampshire / US, from HCRIS.
brfss_summary.csv	BRFSS chronic-disease and access gaps, Vermont vs. peer mean, pre/post the 2018 boundary.
county_dose_response.json	County-level dose-response: change in Medicare spending regressed on change in ACO penetration across 2,702 counties, with the pre-trend placebo and the suppression sensitivity. Resolves far smaller effects than the statewide design and finds none — but the regressor is endogenous, so this is a precise association, not an identified zero.
mde.json	Minimum detectable effect for the statewide synthetic control: the spending gap Vermont would have needed to reach each placebo rank, in dollars, as a share of its spending level, and converted to a per-attributed-life effect. Quoted on the post-period RMSPE scale, the same scale the test uses. The design falls short by roughly a factor of two, not by an order of magnitude.
scm_diagnostics.json	Diagnostics from the level-matched fit. Its ranking statistics, gaps and Medicare Advantage robustness runs are on a different specification from the published ones — use the manifest and `ma_robustness.json` for those. The Medicare Advantage enrolment series by year is the part not available elsewhere.
overdose_vt_index_ci.csv	Vermont overdose deaths indexed to 2015 with Poisson 95% bands — the uncertainty ribbon in figure 3.
suicide_vt_peer_gap.csv	Vermont vs. six-peer age-adjusted suicide rates, 2005–2024, with the WONDER-to-MIOV seam marked. Also consumed by figure 3.
margins_extended.csv	Hospital total and operating margins, five states, 2018–2023, revenue-weighted from HCRIS — the margins table. See the operating-margin caveat in the methodology note before using the operating column across states.
margins_extended_meanofratios.csv	The same margins computed as an unweighted mean of hospital-level ratios rather than revenue-weighted, as a sensitivity check on the aggregation choice.
settlements_attribution.csv	OneCare settlement history and attributed lives by payer and performance year, as far as the public record allows — several GMCB cells are image exhibits and remain incomplete.
aco_penetration_states.csv	Share of Medicare fee-for-service beneficiaries assigned to a Shared Savings Program ACO, by state and year 2016–2019, with the suppressed-cell bounds. The pre-treatment attenuation measure.
aco_penetration.json	Summary of the same: Vermont vs. synthetic Vermont by year, the Next Generation adjustment for donor states, and the 2017–2020 continuity check on one denominator.
aco_penetration_backcast.csv	The same measure back-extended to 2014–2015 from ACO-level results files, with the per-state share of each estimate that relied on the fallback split.
aco_penetration_backcast.json	The back-cast summary, the rejected FFS-proportional rule and why, and the out-of-sample holdout scores (2016 shares → 2017 prediction → 2017 truth).

Public inputs

File	Contents & source
cbp_6211_states.csv	Physician-office (NAICS 6211) establishments and employment by state and year, 2014–2023. Census County Business Patterns.
cbp_6211_counties_vt_nh.csv	Physician-office (NAICS 6211) establishments by county and year, Vermont and New Hampshire, 2014–2023. Census County Business Patterns API.
overdose_state_year.csv	Overdose deaths by state and year, 2015–2025. CDC VSRR provisional counts.
miov_suicide_od_state.csv	Suicide and overdose age-adjusted rates per 100k by state and year, 2019–2024. CDC MIOV.
rand51_state_relative_prices.csv	Commercial prices as a percent of Medicare by state, 2022. RAND Hospital Price Transparency 5.1.

Reproducing the heavy inputs. The synthetic control is fit on the CMS Medicare Geographic Variation PUF; the margins come from HCRIS; the consolidation series from the CBP archives. Those raw files are large and are not mirrored here — they are reproducible from the source links in the methodology note. The derived series above are what the figures actually read.

‹ Back to the accounting Methodology ›