Data coverage

At a glance

From filing to benchmark

A return lists everyone who served the organization, including unpaid board members. Here is each step from what the IRS publishes to the rows our benchmark model is trained on, with what each step removes.

1. What we store

   From each Form 990 we store the people in Part VII, Section A with reported pay of at least $10,000, at filers with revenue of at least $25,000. From each Form 990-EZ we store the people in Part IV with pay of at least $10,000, at filers with revenue of at least $25,000. From Schedule J we store every listed person, at filers with revenue of at least $500,000.

2. Everyone listed on a return, including unpaid board members: not counted. We never store a person below the pay floor, so no stored record holds this number.

3. 3,666,710

   Part VII and 990-EZ person rows stored
4. 1,750,688

   Schedule J person rows stored (the same people usually appear in Part VII too, so the two are never added together)

5. Cleaning

   The stored rows become one compensation record per person and return. We remove people reporting part-time hours, people marked as former, unnamed rows, and board or unclassified titles; keep a person listed in both Part VII and Schedule J once; keep only the latest of an amended or superseded return; and count pay reported by both an organization and a related organization once, at the main employer. Our refresh log records how many rows each of these steps removes, but no stored record this page reads does, so those counts are not shown here.

6. 2,290,751

   Compensation records
7. 2,249,074

   Compensation records that show pay above $0

8. The benchmark model was fitted on a later build of the compensation records than the one counted above. Because they are different builds, we do not subtract one count from the other; the steps below start from the model's own build.

9. Training the benchmark model

   The model is trained on records with pay above $0 and positive organization revenue, then on the rows that pass each rule below.

10. 2,249,509

    Records in the model's build with pay above $0 and positive organization revenue
11. 2,248,413

    After removing Schedule J rows from returns that list no one in Part VII

    1,096 removed
12. 1,679,800

    After the training rules: the filing organization itself paid no one; data flags (no base pay, a retirement payout spike, unusually high other pay); roles the model does not train on (board, former, unclassified); part-year titles (interim, acting, partial year); and one top executive per return

    568,613 removed
13. 1,661,671

    Rows the benchmark model is trained on, after removing rows paid mainly by related organizations rather than the filer, and rows at filers that act as a payroll entity

    18,129 removed
14. Rows failing each training rule. A row can fail more than one rule, so these counts overlap and do not add up to the rows removed.7 rows

    Rows failing each training rule. A row can fail more than one rule, so these counts overlap and do not add up to the rows removed.

    Rule	Training group	Checking group	Total
    The filing organization itself paid no one	128,918	22,566	151,484
    Data flags (no base pay, retirement payout spike, unusually high other pay)	130,251	22,629	152,880
    Roles the model does not train on (board, former, unclassified)	349,604	61,499	411,103
    Part-year titles (interim, acting, partial year)	57,810	9,558	67,368
    More than one top executive on the return	26,653	4,528	31,181
    Filer acts as a payroll entity	5,297	720	6,017
    Paid mainly by related organizations	147,319	26,150	173,469

15. Show the table2 rows

    The model is trained on one group of organizations and checked on a separate group it never saw.

    Group	Rows
    Training group	1,413,292
    Checking group	248,379

Filings by form and fiscal year

Fiscal year is the year a return's fiscal period ends. The most recent years are still arriving. We treat a year as complete once it holds at least 60% as many records as the year before; the latest complete fiscal year is 2025.

Form 990-EZ returns below the revenue floor: opened, but no stored record counts them by form, so they are not in these counts.

Coverage

The comparables roster is a separate table, drawn from the same filings, that lists individual organizations and their reported pay.

Sources

IRS Form 990 e-file releases. We take in the IRS's public e-file XML releases of Form 990 and Form 990-EZ returns.

Release dates: neither the date the IRS published each release nor the date we took it in is kept in a stored record.

BLS Occupational Employment and Wage Statistics (OEWS). Our workforce benchmarks start from the OEWS release named below.

What we do not use, and why. Form 990-PF is filed by private foundations, a different population on a different return. Form 990-T reports unrelated business income tax, not compensation. Form 990-N, the e-Postcard for the smallest organizations, reports no compensation and is not in the e-file XML releases.