Methods

Where the figures come from, how they were checked, and how each measure is calculated.

This document says where every figure in the dataset comes from, how it was checked, and how each measure is calculated. It describes filings, reports and calculations. It makes no statement about why a filer's figures are what they are.

1. Sources

Rhode Island's Defending Affordable Prescription Drug Costs Act (R.I. Gen. Laws chapter 5-19.3) requires each 340B covered entity in the state to report yearly to the Auditor General, who posts the reports. The 14 reports for calendar year 2025 were retrieved from the Auditor General's website on 11 November 2026.

Each PDF is archived unchanged in sources/ri/2025/. manifest.csv records, for each file, the address it was retrieved from, the time, its size, its page count, whether it is a scan, and its SHA-256. The build stops if a PDF no longer matches its SHA-256.

The text each PDF carries is saved beside it in sources/ri/2025/text/, extracted with Poppler's pdftotext (tools/extract_text.py). One thing is changed in that text: an email address is replaced with a marker. The PDFs themselves are unchanged.

All 14 use state Form RIOAG-340B (Revised January 2026): a cover page and Schedules A to I. Thirteen carry a text layer. One (Wood River Health Services) is a scan.

Vermont. The Green Mountain Care Board posts, for each of 13 hospitals, two documents for fiscal year 2025: a one-page "FY2025 340B Program Financial Reporting Detail", marked "As Submitted to the Green Mountain Care Board in 2026", and a free-text "340B Narrative Reporting". The state page that lists them is <https://gmcboard.vermont.gov/node/24841>; each document has its own page at https://gmcboard.vermont.gov/document/<name>. The 26 PDFs were retrieved on 12 November 2026 (UTC; each time is in the manifest). They are archived unchanged in sources/vt/2025/. That manifest.csv has the columns Rhode Island's has, and two more: document_kind (financial or narrative) and cover_page_url. The build stops if a PDF, narratives included, no longer matches its SHA-256.

The 13 financial pages have a text layer; the text is saved beside them in sources/vt/2025/text/, extracted the same way. The narratives' text is not extracted or kept. The page prints this notice, quoted exactly: "NOTE: This is draft and may be adjusted for accuracy in the final GMCB dashboard publication." Every Vermont filing therefore carries the status draft in filings.csv. A later publication by the Board may differ; this dataset holds what the pages printed on the day they were retrieved.

The Vermont page prints no corporate address, 340B ID numbers, recertification date or reporting basis. Those columns are blank for Vermont; nothing is filled in from another source.

Minnesota. Minnesota is a different kind of source. Minn. Stat. 62J.461 has the Department of Health (MDH) collect data from covered entities and report on it yearly to the legislature. The entities' own reports to MDH are not public. What is public, and what this dataset holds, is what MDH printed in its report: "340B Covered Entity Report: Report to the Minnesota Legislature, 2025", cover date February 27, 2026, 51 pages, retrieved on 12 November 2026 (UTC; the time is in the manifest) from <https://www.health.mn.gov/data/340b/docs/2025report.pdf>. It is archived unchanged in sources/mn/2024/, with a manifest.csv of the same layout as Rhode Island's. The build stops if the PDF no longer matches its SHA-256.

The report is titled 2025 and covers the 2024 data year. The dataset's reporting_year is the data year, 2024; the title says 2025. The report's cover letter says that "reported data are not audited nor otherwise verifiable". So every Minnesota figure is a figure the State published about an entity, one per entity (its net 340B revenue, and that figure as a percentage of the statewide total), not a figure the entity filed. filings.csv marks this with record_kind (section 7).

Only Appendix 6 (PDF pages 42 to 45) is transcribed. The text layer of those four pages is saved in sources/mn/2024/text/. The other 47 pages are left empty in that file, because the report names individual people and offices on pages the dataset does not use (tools/extract_text.py --pages 42-45). The PDF is unchanged.

The Minnesota report prints no corporate address, 340B ID numbers, recertification date, reporting basis, status or period in the tables. Those columns are blank for Minnesota.

Idaho. Idaho is a third kind of source. Idaho Code § 41-351 has each covered entity report yearly to the State, keeps those reports confidential, and has the State Controller's Office publish a report that adds them together before November 15 of each year. What is public, and what this dataset holds, is that report: "Idaho 340B Drug Pricing Program, Inaugural Report, Calendar Year 2025", seven pages, linked as "340B Program Reporting" on the Transparent Idaho website (<https://transparent.idaho.gov/featured-resources>) and retrieved on 15 November 2026 (UTC; the time is in the manifest) from the address in sources/id/2025/manifest.csv. It is archived unchanged in sources/id/2025/. The report covers January 1 through December 31, 2025 and says it adds together 62 submissions. It names no covered entity, so the dataset has no Idaho row for any single entity.

The text layer of all seven pages is saved in sources/id/2025/text/. The report prints the State Controller's Office telephone number, fax number and e-mail address; in the saved text each is replaced with a marker (tools/extract_text.py). The PDF is unchanged.

2. What is transcribed

Every printed row of Schedules A to I, one row of data per printed row. For each row the dataset keeps:

  • the text as filed (labels, names, the drug code cell, the amount, the claim count), copied without correction. A name the form cut off stays cut off. Spelling stays as filed;
  • a parsed amount, at the scale filed (some filers report whole dollars, others cents);
  • the PDF page the row is printed on;
  • how the row was checked.

Rules for amounts:

2. What is transcribed: Rules for amounts
Printed in the amount cellamount_as_filedamount
A figurethe figure as printedthe same number
A figure in parenthesesas printednegative
A lone dash-0
N/AN/Ablank
Nothingblankblank

A blank is never treated as zero.

What is not transcribed:

  • The signer's name, title and email address on Schedule I. They are in the PDF.
  • Pages a filer added to the form: cover letters, notes, appendices and supplemental listings. Each is listed in data/attachments.csv with its page range, so a reader knows it exists and where to find it.
  • Notes a filer printed on a schedule page are kept in the note column of the row they sit beside, starting Printed note:.

Vermont. Each page is transcribed in full: the name printed at the top, the title, the submission line and the draft notice (in filings.csv, see DATA_DICTIONARY.md), and the 12 figure rows in printed order. The page prints no line numbers and no schedule letters. So schedule and line hold keys this dataset assigned (for example payments and total), and label_as_filed holds the label the page prints, unchanged. The keys are listed in DATA_DICTIONARY.md. They are ours: the state does not use them.

Further rules for Vermont's amounts, in addition to the table above:

2. What is transcribed: Further rules for Vermont's amounts, in addition to the table above
Printed in the amount cellamount_as_filedamount
$764,691 (no cents)764,691764691, at the scale filed
($397,702)(397,702)-397702
-33.30%-33.30%-33.30, with unit percent
N/RN/Rblank

Every row has a unit (usd, percent or none), fixed by the form, so a percentage is never mixed with dollars. N/R is kept as printed and is not read as zero. The state's text below the figures, where a page has any, is kept word for word as a page_note row. One page has such a paragraph (Northwestern Medical Center); it is carried unchanged and not interpreted.

Not transcribed for Vermont: the narratives. Each is listed in data/attachments.csv with its address, SHA-256 and page count, kind narrative. They are free text, not a form. This dataset holds no person's name, email address or phone number.

Minnesota. Appendix Table 6.1 prints 26 rows, one per covered entity. Each printed row is one record (mn-2024-<key>, with a key this dataset assigns from the name) and has two figure rows, both on the page the row is printed on: table_6_1 / net_revenue (the column "Net 340B Revenue ($)", unit usd) and table_6_1 / pct_of_statewide (the column "% of Statewide Net 340B Revenue", unit percent). In both, label_as_filed is the entity name as printed and description_as_filed the column heading as printed. A name, type or city that wraps across lines in the table is joined with single spaces; nothing else is changed (the apostrophe in "Children’s Minnesota" is the one the page prints, and differs from the other apostrophes). The printed entity type and city are kept as printed in two new columns of entities.csv (entity_type_as_filed, city_as_filed). "As filed" means, for Minnesota, as printed in the State's report.

The existing entity_type column is hospital when the printed type contains the word Hospital (Disproportionate Share Hospital, Sole Community Hospital, Critical Access Hospital, Children's Hospital) and other for any other printed type (one row prints "CHD Treatment Center"). Nothing is read into an abbreviation. The build refuses a Minnesota record whose entity_type does not follow its printed type.

Appendix Table 6.2 prints six rows ("Top 25, unadjusted", "Remaining 156, unadjusted", "Total, unadjusted", and the same three "adjusted"), each with "Net 340B Revenue ($)" and "% of Total". They are kept as 12 figure rows under one statewide record, mn-2024-statewide, whose entity mn-statewide has entity_type state_total and is not a covered entity. The table starts on page 44 and the adjusted rows are on page 45. A "Total" row is row_kind total. The statewide record also holds the report's running text around the two tables as six page_note rows, word for word: the paragraph above Table 6.1; the title of Table 6.1; the "Source:" line under Table 6.1; the paragraph above Table 6.2; the title of Table 6.2; and the "Source:" line under Table 6.2. The two "Source:" lines differ (one names 2023 and 2024 data, the other 2024 data), and the second is printed without a final full stop; both are kept as printed. A page note is not interpreted, except that the rule in section 4 reads the number of entities from the first.

Amounts: 334,683,472 is 334683472; 26.1% is 26.1 with unit percent; 100% is 100. A percentage keeps its printed precision. The printed order of the table is kept nowhere in the data: each record is its own file, and published files are listed by filing_id, which is alphabetical. The report's order is not a field. Nothing in the dataset ranks the rows, and nothing calls a row a place in an order. The words "Top 25" and "Remaining 156" are labels as printed.

Not transcribed for Minnesota: the rest of the report, including its cover letter, its main text and its other appendices. Where the report's tables print an entity's name, that is the only name the dataset holds. This dataset holds no person's name, email address or phone number.

Idaho. The whole report is one record, id-2025-statewide (record_kind is state_report_totals, entity_type is state_total). Every cell the report prints in a table, a highlight tile or its chart is one row, under schedule and line keys this dataset assigns (DATA_DICTIONARY.md lists them): the four highlight tiles, the four labels of the chart, the submissions table, the four tables by payer type (acquisition cost, payments received, payments to contract pharmacies, number of claims) and the table of how savings were used. label_as_filed is the row's printed label and description_as_filed the column heading. Each printed percentage is a row of its own.

Three things are kept in a particular way, each said on the row. The two money tiles print rounded figures with a letter ($812.9M, $430.6M); they are kept as text in answer_as_filed, with no amount, so the dataset never turns a rounded figure into dollars. One cell of the submissions table prints a count and a percentage together, 62 (100%); it is two rows, the count and the percentage, and each row's note gives the cell as printed. The words in the Notes column of that table are rows with no figure.

Sentences of the report that state a figure, the chart's title and the labels of its value axis are page_note rows, exactly as printed (a wrapped line is joined with a space), so that every figure printed on a page is carried by a row. The period the report states is in period_as_filed. Not transcribed: the rest of the running text (the description of the federal program and of the statute, and the list of entity types).

3. How each figure was checked

Three separate checks stand behind every published figure. A row that has not passed them cannot be published: the build refuses it.

Check 1: the PDF's own text. filings340b textcheck reads the text layer of each PDF page and asks two questions. Is every transcribed amount and claim count printed on the page it cites? Is every amount printed on the page carried by a transcribed row? The first catches a mistyped figure; the second catches a row left out.

Check 2: an independent read of the page image. A second reader is shown only pictures of the pages, with no access to the PDF's text or to the first transcription, and writes down every amount, claim count and yes/no mark it sees. The result is kept in verification/. A row is marked as checked only when this read agrees with the transcription, figure for figure as printed.

Check 3: the sums. filings340b tieout runs the rules in section 4.

Both reads were made by software (a language model), not by a person, and each was run as a separate job. The second reader's instructions and file access were restricted to the page images, and the record of what each job opened was reviewed afterward. Anyone can repeat either check: the PDFs, the text, both reads and the code are all in this repository.

The verification column records the outcome for each row:

3. How each figure was checked: The verification column records the outcome for each row
ValueMeaning
text_and_imageThe transcription matches the PDF's text and the independent read of the page image
two_image_readsThe page has no usable text (a scan, or a figure the text layer leaves out); two independent reads of the page image agree
adjudicatedThe two reads differed; a further look at the page image settled it, and the row's note says what was seen
unreadableThe page cannot be read with confidence; the value is left blank

In the Rhode Island 2025 release the two reads agreed on every one of the 2,989 rows. 2,887 are marked text_and_image and 102 two_image_reads (the 101 rows of the scanned filing, and one claim count that a PDF's text layer leaves out). No row needed adjudication and none is unreadable.

In the Vermont 2025 release a separate reader was given only the 13 page images and wrote down every figure row it saw, in printed order, with the label and the value as printed (tools/compare_vt_image_read.py keeps a normalised copy in verification/vt/2025/). The comparison for Vermont is stricter than the shared one: it also counts a parenthesis, a minus sign and a % sign, so a figure read as positive where the page prints it in parentheses would be a difference. The two reads agreed on all 157 rows (13 times 12 figure rows, and the one page_note row, whose text was matched with the reader's notes). All 157 are marked text_and_image. None was adjudicated and none is unreadable.

In the Minnesota 2024 release a separate reader was given only the page images of PDF pages 42 to 45, with no access to the PDF's text or to the transcription, and wrote down every cell of Table 6.1 (name, type, city, net revenue, percent) and Table 6.2 (dimension, net revenue, percent), with the running text around the tables. The image of page 45, where Table 6.2 continues, was read by a separate reader in a separate job. tools/compare_mn_image_read.py keeps a normalised copy of both reads in verification/mn/2024/ and compares every name, type, city and figure with the transcription exactly as printed, apart from the $ sign and runs of spaces (stricter than the shared comparison: a % sign, a comma or an apostrophe that differs is a difference). A record whose name, type or city differs from the read, or a figure the reader marked ?, would stay unverified and unbuilt. The two reads agreed on all 26 Table 6.1 rows (name, type, city, both figures), on all six Table 6.2 rows (both figures; three rows on page 44 and three on page 45), and the six page notes were found word for word in the reader's notes. All 70 rows are marked text_and_image. None was adjudicated and none is unreadable. For Minnesota, filings340b textcheck pools the records that are rows of one PDF page: a figure printed on page 43 counts as carried when any record of the report carries it.

In the Idaho 2025 release a separate reader was given only the pictures of the seven pages, with no access to the PDF's text or to the transcription, and wrote down every cell of every table, tile and chart, with its row label and column heading, and the running text of each page. tools/compare_id_image_read.py pairs each row with the reader's cell on the same page with the same label and heading, compares the two exactly as printed apart from the $ sign and spaces, lists any cell the reader wrote that no row carries, and keeps a normalised copy in verification/id/2025/ beside the reader's own cell file (id-2025-statewide.second-read.csv). A page_note row counts as agreeing only when its exact text is in the reader's running text, is the heading the reader gave a chart, or is the reader's axis labels in order. The two reads agreed on all 105 rows (90 cells and 15 page_note rows), and the reader wrote no cell that is not carried. All 105 are marked text_and_image. None was adjudicated and none is unreadable.

Limits of these checks. Check 2 covers figures and marks: amounts, claim counts, and Schedule I answers. Names and labels come from the PDF's text, and for the scanned filing from one read of the image. Where a drug name is clipped by the form's cell, the dataset keeps the text the PDF holds.

4. Sums and comparisons

Each rule sets two figures from one filing side by side. The basis column says why.

form: the form itself says the two figures agree.

4. Sums and comparisons: form: the form itself says the two figures agree
RuleWhat it compares
A6=A2-A1-A3-A4-A5Schedule A line 6 with line 2 minus lines 1, 3, 4 and 5 (the form's net revenue calculation)
B.rows=total … H.rows=totalThe rows of a schedule with the total printed on it
B.total=A1, C.total=A2, D.total=A3, E.total=A4, F.total=A5A schedule's total with the Schedule A line the form says it supports
C.claims=total, G.claims=totalThe claim counts on a schedule with the claim total printed on it

comparison: the form does not say the two figures agree. They describe related things and are shown together.

4. Sums and comparisons
RuleWhat it compares
H.total=A6The Schedule H total (programs supported by 340B net revenue) with Schedule A line 6 (net 340B revenue)
C.claims=G.claimsThe claim total on Schedule G (drugs) with the claim total on Schedule C (reimbursements)

A rule is applied only when the filing prints both figures.

Vermont. Four rules, run on every Vermont filing. A cell that prints N/R has no figure and adds nothing to a sum.

form: the page itself says the figures agree.

4. Sums and comparisons: form: the page itself says the figures agree
RuleWhat it compares
payments.rows=totalThe four payment rows (by distribution channel) with Total 340B Payments Received
expenses.rows=totalThe three expense rows with Total 340B Expenses
net_revenue=payments-acquisition-expenses340B Net Revenue with Total 340B Payments Received minus 340B Prescription Drug Acquisition Costs minus Total 340B Expenses (the page prints that calculation under 340B Net Revenue)

comparison: the page does not say how the figure is calculated.

4. Sums and comparisons: comparison: the page does not say how the figure is calculated
RuleWhat it compares
operating_margin=net_revenue/payments340B Operating Margin with 340B Net Revenue divided by Total 340B Payments Received, times 100, rounded half up to the number of decimal places the page prints the margin to. The difference is in percentage points (unit percent)

When the two figures differ, nothing is changed. The difference is listed in data/discrepancies.csv with both figures, and a note that states what the filing prints. The build stops if a difference has no note, and stops if a note is left behind for a rule that now holds. The Rhode Island 2025 data lists 22 differences across 13 filings: eight under a form rule (five of them of 1.00 or less) and 14 under a comparison. The Vermont 2025 data lists four, in three filings: two under a form rule of 1.00 each (Porter Medical Center, University of Vermont Medical Center), one under a form rule of 80,000.00 (Gifford Medical Center, net revenue), and one comparison (Gifford Medical Center, operating margin). The other ten Vermont filings, and each of these three on its other rules, hold.

Minnesota. Rules compare figures across the report's records, so each is attached to the record it describes: the entity rows carry the percentage comparison, and the statewide record carries the rest. A rule is applied only when the report prints both figures.

form: the report's text states that the two figures agree.

4. Sums and comparisons: form: the report's text states that the two figures agree
RuleWhat it compares
table_6_1.rows=stated_entitiesThe number of rows Table 6.1 prints beside the number of Covered Entities the paragraph above it states (unit count)
table_6_1.sum=table_6_2.top_unadjustedThe Net 340B Revenue rows of Table 6.1 added together beside Table 6.2's "Top 25, unadjusted"
table_6_2.unadjusted_rows=total, table_6_2.adjusted_rows=totalThe named group plus the remaining entities beside the total, for unadjusted and for adjusted

comparison: the report does not say how the figure is calculated.

4. Sums and comparisons: comparison: the report does not say how the figure is calculated
RuleWhat it compares
pct_of_statewide=net_revenue/total_unadjustedOn each entity row: the printed percentage beside its net revenue divided by Table 6.2's "Total, unadjusted", times 100, rounded half up to the number of decimal places the report prints. The report does not say which total its percentages use; this is one comparison, not a statement of how they were made. The difference is in percentage points
table_6_2.<row>.pct_of_total=net_revenue/total_<unadjusted or adjusted>On the statewide record, for each of its six rows: the printed % of Total beside the row's net revenue divided by the total of its own kind (unadjusted rows by Total, unadjusted; adjusted rows by Total, adjusted), at the printed precision

The Minnesota 2024 data lists seven differences: three under a form rule, all on the statewide record (Table 6.1 prints 26 rows where the paragraph above it states 25; the Table 6.1 rows add up to 1,152,403,308 where Table 6.2 prints 1,152,403,307; the adjusted rows add up to 1,340,270,965 where Table 6.2 prints 1,340,270,964), and four under the entity percentage comparison (Abbott Northwestern Hospital, CentraCare - Rice Memorial Hospital, M Health Fairview - University of Minnesota Medical Center and Regions Hospital, each 0.1 percentage points below the comparison). The other 22 entity rows hold, and so do the unadjusted rows of Table 6.2 and all six of its percentage comparisons.

Idaho. Every rule is on the one statewide record.

4. Sums and comparisons: Idaho. Every rule is on the one statewide record
RuleBasisWhat is compared
<table>.rows=totalformFor each of the four payer tables and the savings table: the rows above the TOTAL row added together, beside the TOTAL row
<table>.<row>_pct=<row>/totalcomparisonEach printed percentage beside the row's figure divided by the table's TOTAL, times 100, at the printed precision
submissions.complete_pct=complete/receivedcomparisonThe percentage printed beside Submissions with complete status, beside that count divided by Total submissions received
key_highlights.total_claims=claims.total and key_highlights.covered_entities=submissions.receivedcomparisonA count tile beside the table figure of the same name
key_highlights.payments_received=payments_received.total and key_highlights.savings_to_community=savings_use.totalcomparisonA money tile, a rounded figure in millions, beside the table's TOTAL rounded the same way
key_financial_metrics.<label>=<table>.totalcomparisonEach of the chart's four labels beside the table TOTAL of the same name, to the whole dollar
savings_use.total=payments_received.total-acquisition_cost.totalcomparisonThe TOTAL of 340B Savings Utilization beside payments received less acquisition cost, the difference the report's summary calls the gross program margin. The report does not say the two agree

The percentage columns are not added up: each percentage is printed to one decimal.

The Idaho 2025 data lists three differences. Two are under a form rule: the six payer rows of Number of Claims by Payer Type add up to 1,916,602 where the TOTAL row prints 1,916,995, and the nine rows of 340B Savings Utilization add up to 430,646,947.35 where the TOTAL row prints 430,646,950.35. One is under a comparison: the savings TOTAL, 430,646,950.35, beside payments received less acquisition cost, 413,709,636.77. The other 50 rules hold.

Schedule G's amount total is not compared with Schedule A line 1. The form does not say they agree, and the filings describe Schedule G's amounts in more than one way.

All differences, by filing

36 differences across 22 filings, reports and report rows: 16 under a form rule and 20 under a comparison rule. Nothing was changed; both figures are kept as filed.

Every difference found between two figures in one filing or report
Filing or report rowRuleBasisExpectedFiledDifferenceNote
Abbott Northwestern Hospitalpct_of_statewide=net_revenue/total_unadjustedcomparison12.1%12.0%-0.1 percentage pointsThe report does not say which total its percentages use. Table 6.1 prints 12.0% for this row. This row's Net 340B Revenue of 153,985,063 divided by Table 6.2's Total, unadjusted, 1,276,243,781, times 100, is 12.1%, to one decimal.
Blackstone Valley Community Health CareG.rows=totalform3,195,260.133,195,260.140.01The rows of Schedule G as filed add up to 3,195,260.13; the total printed on the schedule is 3,195,260.14.
CentraCare - Rice Memorial Hospitalpct_of_statewide=net_revenue/total_unadjustedcomparison0.6%0.5%-0.1 percentage pointsThe report does not say which total its percentages use. Table 6.1 prints 0.5% for this row. This row's Net 340B Revenue of 7,029,723 divided by Table 6.2's Total, unadjusted, 1,276,243,781, times 100, is 0.6%, to one decimal.
Comprehensive Community Action, Inc.B.total=A1form250,279.19250,270.19-9.00Schedule B prints a total of 250,270.19. Schedule A line 1, which Schedule B says it supports, prints 250,279.19.
Comprehensive Community Action, Inc.C.claims=G.claimscomparison3,1354,9601,825Schedule C prints 3,135 claims in total and Schedule G prints 4,960. The form does not say these two counts agree; they are shown side by side.
East Bay Community Action ProgramC.claims=G.claimscomparison8,3328,35220Schedule C prints 8,332 claims in total and Schedule G prints 8,352. The form does not say these two counts agree; they are shown side by side.
Gifford Medical Centernet_revenue=payments-acquisition-expensesform2,800,632.002,880,632.0080,000.00The page prints Total 340B Payments Received of 3,159,915.00, 340B Prescription Drug Acquisition Costs of 163,009.00 and Total 340B Expenses of 196,274.00. By the calculation the page prints under 340B Net Revenue (payments received minus acquisition costs minus expenses) these give 2,800,632.00. The page prints 340B Net Revenue of 2,880,632.00.
Gifford Medical Centeroperating_margin=net_revenue/paymentscomparison91.16%88.63%-2.53 percentage pointsThe page prints 340B Net Revenue of 2,880,632.00 and Total 340B Payments Received of 3,159,915.00; the first divided by the second is 91.16%. The page prints a 340B Operating Margin of 88.63%. The page does not say how the Operating Margin is calculated; the two figures are shown side by side.
Idaho: statewide totalsclaims.rows=totalform1,916,6021,916,995393The six payer rows of Number of Claims by Payer Type add up to 1,916,602. The table's TOTAL row prints 1,916,995, the figure the report also prints in its Total Claims Dispensed tile and in the paragraph above the table.
Idaho: statewide totalssavings_use.rows=totalform430,646,947.35430,646,950.353.00The nine rows of 340B Savings Utilization add up to 430,646,947.35. The table's TOTAL row prints 430,646,950.35, the figure the report also prints in the paragraph above the table.
Idaho: statewide totalssavings_use.total=payments_received.total-acquisition_cost.totalcomparison413,709,636.77430,646,950.3516,937,313.58The report does not say these two figures agree; they are shown side by side. Its summary describes payments received less acquisition cost as a gross program margin of approximately $413.7 million; the two TOTAL rows give 812,855,944.90 minus 399,146,308.13, which is 413,709,636.77. The TOTAL row of 340B Savings Utilization prints 430,646,950.35.
Kent County Memorial HospitalH.total=A6comparison37,941,584.0042,752,942.004,811,358.00Schedule H prints a total of 42,752,942.00. Schedule A line 6 prints 37,941,584.00. The form does not say these two figures agree; they are shown side by side.
Landmark Medical centerF.total=A5form150,994.00150,944.00-50.00Schedule F prints a total of 150,944.00. Schedule A line 5, which Schedule F says it supports, prints 150,994.00.
Landmark Medical centerH.total=A6comparison4,478,236.245,817,246.981,339,010.74Schedule H prints a total of 5,817,246.98. Schedule A line 6 prints 4,478,236.24. The form does not say these two figures agree; they are shown side by side.
M Health Fairview - University of Minnesota Medical Centerpct_of_statewide=net_revenue/total_unadjustedcomparison26.2%26.1%-0.1 percentage pointsThe report does not say which total its percentages use. Table 6.1 prints 26.1% for this row. This row's Net 340B Revenue of 334,683,472 divided by Table 6.2's Total, unadjusted, 1,276,243,781, times 100, is 26.2%, to one decimal.
Minnesota: statewide totalstable_6_1.rows=stated_entitiesform2625-1The paragraph above Table 6.1 states 25 Covered Entities. Table 6.1 prints 26 rows. Table 6.2 labels its first group Top 25.
Minnesota: statewide totalstable_6_1.sum=table_6_2.top_unadjustedform1,152,403,3081,152,403,307-1The 26 Net 340B Revenue rows of Table 6.1 add up to 1,152,403,308. Table 6.2 prints 1,152,403,307 for Top 25, unadjusted.
Minnesota: statewide totalstable_6_2.adjusted_rows=totalform1,340,270,9651,340,270,964-1Table 6.2 prints 1,199,240,413 for Top 25, adjusted and 141,030,552 for Remaining 156, adjusted, which add up to 1,340,270,965. It prints 1,340,270,964 for Total, adjusted.
Porter Medical Centerpayments.rows=totalform9,824,294.009,824,295.001.00The four payment rows as filed add up to 9,824,294.00; the Total 340B Payments Received printed on the page is 9,824,295.00.
Regions Hospitalpct_of_statewide=net_revenue/total_unadjustedcomparison3.8%3.7%-0.1 percentage pointsThe report does not say which total its percentages use. Table 6.1 prints 3.7% for this row. This row's Net 340B Revenue of 47,883,168 divided by Table 6.2's Total, unadjusted, 1,276,243,781, times 100, is 3.8%, to one decimal.
Rhode Island HospitalC.claims=G.claimscomparison441,658982,072540,414Schedule C prints 441,658 claims in total and Schedule G prints 982,072. The form does not say these two counts agree; they are shown side by side.
Rhode Island HospitalH.rows=totalform411,784,983.92411,784,983.930.01The rows of Schedule H as filed add up to 411,784,983.92; the total printed on the schedule is 411,784,983.93.
Rhode Island HospitalH.total=A6comparison398,789,776.02411,784,983.9312,995,207.91Schedule H prints a total of 411,784,983.93. Schedule A line 6 prints 398,789,776.02. The form does not say these two figures agree; they are shown side by side.
The Miriam Hospital Immunology CenterC.claims=G.claimscomparison2,3962,299-97Schedule C prints 2,396 claims in total and Schedule G prints 2,299. The form does not say these two counts agree; they are shown side by side.
The Miriam Hospital Immunology CenterH.total=A6comparison1,211,459.83904,847.42-306,612.41Schedule H prints a total of 904,847.42. Schedule A line 6 prints 1,211,459.83. The form does not say these two figures agree; they are shown side by side.
The Miriam Hospital RISE TB ClinicC.claims=G.claimscomparison2,6362,557-79Schedule C prints 2,636 claims in total and Schedule G prints 2,557. The form does not say these two counts agree; they are shown side by side.
The Miriam Hospital RISE TB ClinicH.total=A6comparison48,725.7340,000.00-8,725.73Schedule H prints a total of 40,000.00. Schedule A line 6 prints 48,725.73. The form does not say these two figures agree; they are shown side by side.
The Miriam Hospital Ryan White Care ProgramC.claims=G.claimscomparison22,29321,414-879Schedule C prints 22,293 claims in total and Schedule G prints 21,414. The form does not say these two counts agree; they are shown side by side.
The Miriam Hospital Ryan White Care ProgramH.total=A6comparison6,830,161.142,890,618.68-3,939,542.46Schedule H prints a total of 2,890,618.68. Schedule A line 6 prints 6,830,161.14. The form does not say these two figures agree; they are shown side by side.
The Providence Community Health Centers, IncG.rows=totalform8,143,697.918,143,697.89-0.02The rows of Schedule G as filed add up to 8,143,697.91; the total printed on the schedule is 8,143,697.89.
Thundermist Health CenterG.rows=totalform14,485,262.6314,485,262.62-0.01The rows of Schedule G as filed add up to 14,485,262.63; the total printed on the schedule is 14,485,262.62.
Thundermist Health CenterH.total=A6comparison10,793,866.7910,895,928102,061.21Schedule H prints a total of 10,895,928. Schedule A line 6 prints 10,793,866.79. The form does not say these two figures agree; they are shown side by side. The filer printed a note beside the Schedule H total; it is carried in that row's note.
University of Vermont Medical Centerexpenses.rows=totalform13,142,032.0013,142,031.00-1.00The three expense rows as filed add up to 13,142,032.00; the Total 340B Expenses printed on the page is 13,142,031.00.
Women & Infants Hospital of RIG.rows=totalform55,141,406.8355,496,300.74354,893.91The rows of Schedule G as filed add up to 55,141,406.83; the total printed on the schedule is 55,496,300.74.
Women & Infants Hospital of RIH.total=A6comparison27,737,405.9841,766,215.0014,028,809.02Schedule H prints a total of 41,766,215.00. Schedule A line 6 prints 27,737,405.98. The form does not say these two figures agree; they are shown side by side.
Wood River Health Services, Inc.G.rows=totalform1,677,9911,677,990-1The rows of Schedule G as filed add up to 1,677,991; the total printed on the schedule is 1,677,990.

5. Measures

5. Measures
MeasureCalculation
gross_marginSchedule A line 2 minus line 1
external_costSchedule A line 3 plus line 4
external_share_of_marginexternal_cost divided by gross_margin, to six decimal places; blank when the margin is not above zero
schedule_e_third_party_administrator, schedule_e_referral_capture, schedule_e_consulting_audit, schedule_e_otherThe sum of the Schedule E rows in that category

Measures use the Schedule A figures as filed, including where a supporting schedule's total differs from them.

Vermont. The same three measures, on Vermont's rows, with the formula text in measures.csv naming them:

5. Measures
MeasureCalculation
gross_marginTotal 340B Payments Received minus 340B Prescription Drug Acquisition Costs
external_costPayments Made to Contract Pharmacies plus Payments Made to Outside Vendors
external_share_of_marginexternal_cost divided by gross_margin, to six decimal places; blank when the margin is not above zero (the same rule as Rhode Island)

Minnesota has no measures. Its tables print no acquisition cost and no payments to pharmacies or vendors, so gross_margin, external_cost and external_share_of_margin do not exist for it, and measures.csv has no Minnesota rows. Net revenue and its share of the statewide figure are the State's published figures, kept in line_items.csv, not measures calculated here.

Idaho has no measures. Its one record is a statewide total and the measures compare covered entities, so measures.csv has no Idaho rows. The report's own totals, and the margin its summary describes, are in line_items.csv and discrepancies.csv as printed.

Internal 340B Program Expenses is not part of external_cost. The Rhode Island Schedule E category measures do not exist for Vermont, which has no such schedule. One Vermont filing (Northwestern Medical Center) has a gross margin below zero, so its share is blank. The measures use the printed Total 340B Payments Received, including where a total differs from its rows. The chart in the README is Rhode Island's by design; Vermont's measures are in data/measures.csv; Minnesota has none.

Why two filers' shares are not directly comparable

A filer that dispenses through its own pharmacy carries the cost of dispensing in its own staffing and expenses. A filer that relies on contract pharmacies pays for dispensing on Schedule A line 3. The same activity therefore appears inside the margin for one filer and as an external cost for another. Line 3 also counts every pharmacy a filer lists on Schedule D, whoever owns it. Filers also differ in reporting basis (cash or accrual, shown in data/filings.csv) and in how they describe the period their figures cover; several print notes about this, which are kept. For these reasons the dataset lists filings alphabetically within health centers and hospitals and does not rank them.

The same holds for Vermont's hospitals: the filings are made by different hospitals in different ways, Vermont's page prints no reporting basis, and every page is marked draft. Nothing here ranks them.

6. Vendor names and categories

normalized_vendor gives one spelling for a company that filers name in different ways, so its rows can be added up. The name as filed is always kept beside it. A name is normalized only when the text as filed names one company. A row that names several companies together, or a category instead of a company, has no normalized name. reference/vendors.csv lists every alias and the reason for any mapping that is not a plain spelling difference.

category is given to Schedule E rows only, and only from the filer's own words on that row: third_party_administrator, referral_capture, consulting_audit, or other when the filer's words name none of these.

7. Rebuilding

filings340b build writes every file in data/ from the archived PDFs' manifest, the transcriptions and the reference files. Each filing names its form (form in its .meta.toml file), and the form decides which rows may appear, their units, and which sums and measures apply; nothing is assumed from the state. Each record also has a record_kind (meta.toml, record_kind): filing for an entity's own filing (Rhode Island, Vermont; the default), state_report_row for a row the State published about one entity (each Minnesota Table 6.1 row), and state_report_totals for the one record that holds the State's totals and running text (mn-2024-statewide, and Idaho's only record, id-2025-statewide). The form says which kinds it allows. It uses no clock and no network, so the same inputs always give the same bytes. filings340b check rebuilds in memory and compares; continuous integration runs it on every change.

8. Versions

The dataset follows semantic versioning. A correction to a published figure is a patch release and is listed in the changelog with the filing, schedule and line. A new state or year is a minor release (0.2.0 added Vermont, 0.3.0 added Minnesota, 0.5.0 added Idaho), and so is a new table (0.6.0 added the state laws). A change to a column's meaning is a major release.

9. Small claim counts

From version 0.4.0, a claim count of 1 to 5 is not republished. The reader sees fewer_than_6 instead. The rule is set in one place, src/filings340b/smallcounts.py, and works on each filing and schedule on its own.

  • A row of kind detail whose claim count is 1, 2, 3, 4 or 5 is small. In data/line_items.csv and in the JSON file, its claims_as_filed and claims are empty and claims_withheld is fewer_than_6. Its drug name, NDC text, amount, page, verification and every other field stay as filed.
  • A count of 0, a blank count and a count of 6 or more are published as filed, and claims_withheld is empty.
  • A small count could be worked out by subtraction from the printed total when a schedule has one small row, or when every small row is a 1 or every small row is a 5. In those cases the detail row with the smallest count of 6 or more is also left out, and its claims_withheld is beside_small. If two rows share that count, the one on the earlier line is used. If the schedule has no such row, the build stops with a message that names the filing and the schedule, and a person decides.
  • If a total row or a below_threshold row has a claim count of 1 to 5, the build stops with a message that names the filing and the schedule, and a person decides. These rows are never left out silently.

The filing's PDF still prints each count, and every row links to its page. Each figure was read from the PDF and checked (section 3) before it was left out. The checks in section 4 publish only sums and printed totals; no single left-out count is published or can be read from a note, a description or a calculated figure.

The rule covers every table file in the repository, not only data/. In transcriptions/ the claims_as_filed cell of a left-out row is empty, and in verification/ the claims cell of the same row is empty. Each of those files has a last column, claims_withheld, that holds the same reason as the published files. The PDFs in sources/ and the page text beside them are the State's public records and are not changed.

The left-out counts are kept in private/small-counts.csv: one row for each, with the count as transcribed and the count the second reader wrote. That file is never committed (it is in .gitignore). The checks run in one of two ways, and each of check, tieout and textcheck prints a line that says which:

  • With the private file, the counts are put back in memory. Every sum is recomputed, the text check and the image-read comparison read every cell, and the rule and its subtraction check run on the real counts. The program stops if the file and the claims_withheld marks do not match one to one, or if a count in the file is not 1 to 5 on a fewer_than_6 row. It also stops if verification/<state>/<year>/claim-sums.csv is not what the counts give.
  • Without it (a clone of the public repository, and continuous integration), the build and check give the same data/ byte for byte. The sum of a schedule's claim counts is read from claim-sums.csv, and the text check and the image-read comparison skip the left-out cells and say how many they skipped. The program stops if a sum in that file cannot match the visible rows.

claim-sums.csv has one row for each filing and schedule that has a left-out count: the sum of the claim counts of its rows, not counting the printed total. The maintainer's run rewrites it with filings340b seal and stops if the committed copy is different, so the committed copy is always one that was checked against the real counts.

After transcribing a new filing with its counts as read, the maintainer runs uv run filings340b seal. It moves the small counts into the private file, empties them in the transcription and the image-read file, and writes claim-sums.csv. Running it again changes nothing. A test walks every CSV file under transcriptions/, verification/ and data/ and fails if a claim cell shows a count of 1 to 5.

In this version the rule leaves out 310 counts as fewer_than_6 and 0 as beside_small. All 310 are Rhode Island Schedule G drug rows (of 2,364 such rows). No Schedule C row has a count of 1 to 5. Every filing with small rows has at least two, and the sum of their counts is more than the number of rows and less than five times that number, so no row is left out under the subtraction case.

10. State laws

From version 0.6.0 the dataset has a table of enacted state laws that name the 340B program. It is kept by hand in laws/laws.csv, one row per enacted act, and published as data/state-laws.csv.

Where a law comes from. Each law was found by searching, then read on an official page: the legislature's own site, the state's code site, or a state office such as a secretary of state. laws/hosts.csv lists every site the table links to, the state it belongs to and who runs it; the build stops at a link to any other site. Lists kept by others were used only to find laws to look for. Nothing in a row is taken from them.

What is saved. tools/save_law_text.py saves the words of each official page under sources/laws/<state>/<law_id>.txt, and a second official page, where one is used, as <law_id>.more.txt. A second page is there for what the first does not print: a bill's history page for its dates, or a code section for its number. source_form says how the first page was read: web page, pdf, or scanned pdf for a scan read by character recognition (one law, Tennessee's 2025 act). In the saved text, curly quotation marks and long dashes are made plain, runs of spaces become one space, the line numbers in the margin of a numbered bill are taken off, and an e-mail address or a telephone number is replaced by a marker. Nothing else is changed. Each row carries the SHA-256 fingerprint of its saved text.

What the build checks. src/filings340b/laws.py stops the build unless, for every row:

  • the saved text names the program (340B, or 42 U.S.C. 256b);
  • quote is in the saved text word for word, and itself names the program;
  • a filled signed_date or effective_date is printed in one of the row's saved texts;
  • each link is on a site listed for that state, and each saved text matches its fingerprint;
  • each kind is one of the seven in laws/kinds.csv.

What is listed. An act is listed when it sets a rule about the 340B program: for drug makers, for insurers and pharmacy benefit managers, for covered entities, for Medicaid, or for a state office. These are left out and named in laws/not-listed.csv, each with its reason: a budget condition whose named years are over; a section since replaced by a listed one; an act that re-enacts a section with the same 340B words as a listed act; an act that names 340B only inside a definition; and a law that came up in the search and whose official text was not read.

Dates and numbers. A date is filled only where a saved page prints it as that date. Where an act takes effect a set number of days after a session ends, or on passage, and the page prints no calendar date, effective_date is blank, limits says how the act takes effect, and source_note says that the page prints no date. A bill number or an act number is blank where the official page does not print it. session_year is the year of the session that passed the act.

The words in a row. summary, binds, limits and enforcement are written by hand from the saved text, in plain words. They describe what the text says. They do not say whether a law is in force today, how a court has read it, or what anyone should do about it. source_note is kept apart from them: it is about the saved pages, not the law, and says what a page does not print or how it was read.

What the table does not cover. Court cases about these laws; bills that did not pass; federal law and guidance; rules made by agencies under these laws. An act amended later is described as enacted, with a note in limits where a saved page shows the later change.

How complete it is. The search covered the fifty states and the District of Columbia. Nothing was found on official pages for thirteen of them (the README names them). A state with no row is a state where none was found. For comparison, the National Conference of State Legislatures wrote in March 2026 that 19 states enacted 22 bills on 340B in 2025; this table has 18 laws from 2025 sessions, in 17 states. The two were not matched bill for bill.

Keeping it current. retrieved is the day the official pages were saved. A page that changes later is not picked up until the law is read again.