How Margin Watch data is built — and how we make sure it’s right

The pipeline

1. Source: the primary file, nothing in between

The Medicare Part B payment-limit data comes from one place: the quarterly ASP pricing files (officially the Medicare Part B Drug Payment Limit Files) published on cms.gov. We hold the complete archive — 87 quarterly files, January 2005 through July 2026— downloaded directly from CMS, public-domain data with no intermediary. No scraped aggregator, no licensed feed, no third-party cleanup we can’t inspect.

2. Parsing two decades of format drift

CMS’s file format changed many times in twenty years — file naming conventions, spreadsheet layouts, header phrasing, even prose legal notes that mention “HCPCS” and “payment limits” in ways that fool naive parsers. Our parser locates the true header row in every era’s layout, rejects blank and non-numeric payment limits rather than coercing them, and refuses to emit a quarter at all if it can’t identify the payment-limit column with confidence. A failed quarter is a loud failure, never a silent gap.

3. Full re-pull every quarter — because CMS rewrites history

CMS restates already-published quarters retroactively: the “January 2026” file on cms.gov today is a December revision of the original. A pipeline that only fetches the newest file slowly diverges from the record. Ours re-downloads the entire archive on every refresh, so restatements are picked up everywhere they land, and pages label figures as publishedper file version — including CMS’s preliminary-vs-final file distinction.

Verification

Cross-validation against an independent dataset

Before first publication (June 2026), we cross-checked the parsed archive against an independently maintained six-quarter ASP dataset from our sister product, CareCost Estimate — 5,356 overlapping (code, quarter) cells. Agreement was 99.93%, and every one of the four mismatches traced to a CMS retroactive restatement, not a parse error.

Known-value spot checks

Each refresh re-verifies landmark values — codes whose payment limits we know from independent records, including structural events like the 2026 ustekinumab biosimilar repricing — and sanity-checks the shape of the archive itself (hundreds of codes must show 15+ years of continuous history, or the run fails).

The product runs on the same math

The paid CareCost Optimizer engine consumes the same reimbursement basis and is held to penny-exact golden tests against an independent reference implementation. The public pages and the product can’t quietly disagree, because they share the data.

Publication standards

Every figure carries its date

Pages state the CMS file version, its posted date, the quarter the limits take effect, and the data vintage of any computed figure. If a number doesn’t have an “as of,” we don’t publish it.

What we will never publish

Margin Watch pages use public-domain sources only: CMS files, FDA records, statutes, and payer policies cited to their primary documents. We do not publish per-payer contracted rates, and we do not maintain live per-payer prior-authorization grids — a stale row there is a denied claim someone acted on. Policy content is explainer and tracker format, dated, with links to the controlling document.

Corrections

Errors are logged, dated, and described on the public corrections & changelog page. Found one? Email hello@carecostoptimizer.com — corrections ship with the next refresh or sooner, and material ones are noted on the affected page.

Citation

Margin Watch tables and statistics are free to cite with attribution to CareCost Optimizer and a link. The underlying CMS data is public domain; the computation, change detection, and framing are ours.

Cadence

CMS posts each quarter’s file roughly three to four weeks before the quarter starts (January, April, July, October). Margin Watch change trackers refresh within days of each release; slower-moving trackers (state sourcing statutes, IRA milestones) update on their source’s own clock, each with a last-verified date on the page.

Modeling details for the paid product — inputs, calculation logic, source-conflict hierarchy, limitations — live on the methodology page. The editorial process is described in our editorial policy.