Field-Data Essay Series

Counting the Killings

The state's count failed. Civilians converged. The total is rising.

The United States does not reliably count the people its police kill; the authoritative national numbers are built by a newspaper, activists, and a volunteer. This page leads with the evidence: every chart below re-derives, at render time, from a cited, hashed open-data release you can download and re-check. The essays and live instruments follow. Systems and design only, no partisan claims.

What the data shows

Finding 01

The count is rising, and the databases disagree by exactly one definition

Both living counts are up roughly 18 percent across the decade, with 2023 and 2024 the highest years on record. The gap between the two lines is not a dispute: Mapping Police Violence starts from the Post's shootings and adds the killings a shooting-only rule cannot see, about 109 people a year. Toggle the rule and watch them appear.

the shaded band is one definitional choice, ~109 people/yr
9001,0001,1001,2001,3002013201520172019202120232013: 1,075 killings, all-killings rule (MPV)2014: 1,039 killings, all-killings rule (MPV)2015: 1,099 killings, all-killings rule (MPV)2016: 1,065 killings, all-killings rule (MPV)2017: 1,091 killings, all-killings rule (MPV)2018: 1,138 killings, all-killings rule (MPV)2019: 1,099 killings, all-killings rule (MPV)2020: 1,161 killings, all-killings rule (MPV)2021: 1,148 killings, all-killings rule (MPV)2022: 1,204 killings, all-killings rule (MPV)2023: 1,248 killings, all-killings rule (MPV)2024: 1,270 killings, all-killings rule (MPV)2015: 995 fatal shootings, shootings-only rule (Post)2016: 959 fatal shootings, shootings-only rule (Post)2017: 984 fatal shootings, shootings-only rule (Post)2018: 992 fatal shootings, shootings-only rule (Post)2019: 993 fatal shootings, shootings-only rule (Post)2020: 1,021 fatal shootings, shootings-only rule (Post)2021: 1,050 fatal shootings, shootings-only rule (Post)2022: 1,097 fatal shootings, shootings-only rule (Post)2023: 1,164 fatal shootings, shootings-only rule (Post)2024: 1,175 fatal shootings, shootings-only rule (Post)2015: 1,146 killings, The Counted (the Guardian), its only two years2016: 1,093 killings, The Counted (the Guardian), its only two years2024, highest on record
All killings (Mapping Police Violence), circlesFatal shootings only (Washington Post), squaresThe Counted (Guardian), 2015-16, diamonds
Interactive chart; the full year-by-year table follows below.

src: WaPo Fatal ForceMapping Police Violenceessay 04, the definition gap

Finding 02

The only federal count is voluntary, and it still can't stay publishable

The FBI's Use-of-Force collection gates its national totals on covering 80 percent of sworn officers. Participation being voluntary, it took six years to clear that bar, in 2024, and fell back below it the next year. A count that depends on volunteering never finishes.

0%25%50%75%100%80% needed to publish a national total2019: 46% of sworn officers covered46%20192022: 53% of sworn officers covered53%20222024: 80% of sworn officers covered (cleared the threshold)80%20242025: 78% of sworn officers covered (fell back below)78%2025publishablefell back

src: USAFacts on the FBI collectionthe FBI collection itselfessay 02, the federal blind spotthe live coverage curve

Finding 03

The government's own death records missed more than half

The deepest undercount is the state's. A peer-reviewed reconstruction in The Lancet found the National Vital Statistics System failed to attribute 55.5 percent of deaths from police violence over 1980-2018, 17,100 of an estimated 30,800, filed under another cause. The misclassification was pervasive for every group and highest for Black Americans, at 59.4 percent, a measurement finding about whose deaths the certificates failed to count.

17,100 deaths filed under another cause (55.5%)13,700 deaths attributed to police violence17,100 missed (55.5%)of 30,800SHARE MISSED, BY GROUP (SAME STUDY)Black non-HispanicBlack non-Hispanic: 5,670 of 9,540 missed (59.4%)59.4%White non-HispanicWhite non-Hispanic: 8,540 of 15,200 missed (56.2%)56.2%Hispanic, any raceHispanic, any race: 2,580 of 5,170 missed (49.9%)49.9%Other non-HispanicOther non-Hispanic: 281 of 861 missed (32.6%)32.6%

src: The Lancet (GBD 2019)essay 06, how many did nobody count

Finding 04

Three independent civilian trackers agree on the same people

Match the lists person by person, from MPV's own linkage columns, and rival projects turn out to know about the same deaths: in 2019, 99.4 percent of MPV's cases were also on the Post's or Fatal Encounters' lists, and no overlap year falls below 98.4 percent. Civil society converged where the state could not.

1,092 of 1,099 also on the Post's or Fatal Encounters' lists7 killings only MPV caught in 201999.4% also on another civilian list (2019)lowest overlap year still agrees on 98.4% of the people

src: MPV linkage, re-derivedFatal EncountersThe Counted (Guardian)essay 06

Finding 05

The rise is where reform isn't

Mapping Police Violence's year-end analysis finds killings declining in the biggest cities, many of which rewrote use-of-force policy over the past decade, while rising in suburban and rural areas policed by thousands of smaller agencies, and the record years arrived while national homicide figures were falling. The national total is the sum of those two curves.

Carried as a cited finding, not a chart: this dataset holds aggregates only, never the record-level rows a geographic split would need. Honesty about what we hold is part of the method.

src: MPV 2024 year-end reportthe trend reading in essay 06

Finding 06

The government tested the civilian method. It worked.

Before any of this was settled, the Bureau of Justice Statistics assessed its own Arrest-Related Deaths program and found it captured, at best, 49 percent of law-enforcement homicides. It suspended the collection in 2014, piloted the open-source media review the civilian trackers use, and the pilot identified 1,348 potential deaths in ten months, an estimated 1,900 for the year. The state's own statisticians validated the method in 2016; the finding sat in a PDF.

49%
of law-enforcement homicides captured by the old federal program, at best
~1,900
arrest-related deaths a year found by the redesigned media-review pilot

src: BJS assessment (NCJ 248543)BJS redesign pilot (NCJ 250112)

Finding 07

An outlier among peer countries, every one of them counting differently

The US rate, 33.5 per 10 million people, runs 3.4 times the next-highest peer. The honest caveat is the series' own thesis: every country counts under its own definition and system, so the comparison is indicative, not harmonized. The US figure derives from MPV's 2019 count of 1,099, the same number in the chart at the top of this page.

United StatesUnited States: 33.5 per 10 million (2019, Mapping Police Violence)33.5CanadaCanada: 9.8 per 10 million (2017, CBC News database)9.8AustraliaAustralia: 8.5 per 10 million (2017-18, National Deaths in Custody Program)8.5NetherlandsNetherlands: 2.3 per 10 million (2019, Public Prosecution Service)2.3New ZealandNew Zealand: 2.1 per 10 million (2018, NZ Police Tactical Options Research Report)2.1GermanyGermany: 1.3 per 10 million (2018, DPA via Deutsche Welle)1.3England & WalesEngland and Wales: 0.5 per 10 million (2019, INQUEST)0.5JapanJapan: 0.2 per 10 million (2018, Axios)0.2IcelandIceland: 0 per 10 million (2018, NBC News)0NorwayNorway: 0 per 10 million (2018, Norwegian Bureau for Investigation of Police Affairs)0

src: Prison Policy Initiative (2020)the release's international table

Finding 08

The rise and the blind spot are the same agencies

The board's findings meet in one place. This dataset carries the federal collection measured two ways in the same year: agencies that report employ 53 percent of the country's officers, but only 37 percent of departments file at all. The big departments report; the missing mass is thousands of small agencies. The rise in finding 05, meanwhile, is concentrated in suburban and rural areas policed by exactly those agencies. Whatever each department's reason for silence, the killings are growing in exactly the agencies that do not report them, which is this series' whole thesis stated as a finding.

share of sworn officers coveredshare of sworn officers covered, 2022: 53%53%share of departments filingshare of departments filing, 2022: 37%37%
2022 · the difference is thousands of small departments

src: FBI collection participation (2022)MPV 2024 year-end reportessay 08, the two whys

Finding 09

The disparity survives every definition

A series about definitional fragility owes special weight to the number no definition can move. Every counter here that publishes rates by race finds Black Americans killed at two to three and a half times the white rate: the Post's shootings-only rates, MPV's all-killings ratio, and the Lancet's forty-year reconstruction, three rules, three eras, one direction. These are population-level measurements, quoted from each instrument's own figures; the essay names the denominator debate honestly and asserts no cause. The counting failure compounds it: the group with the highest rate is also the group whose deaths official records misclassified most (finding 03).

1x parity2x3x4xWashington Post Fatal ForceFatal shootings only · 2015-2024Washington Post Fatal Force (Fatal shootings only, 2015-2024): 2.5x the white rate, 6.1 vs 2.4 per 1,000,000 per year2.5xMapping Police ViolenceAll killings · 2024Mapping Police Violence (All killings, 2024): 2.9x the white rate2.9xGBD 2019 (The Lancet)All police violence, reconstructed · 1980-2019GBD 2019 (The Lancet) (All police violence, reconstructed, 1980-2019): 3.5x the white rate, 0.69 vs 0.2 per 100,000, age-standardized3.5x
each instrument's own published figure · dotted line is parity

src: WaPo databaseMPV 2024 reportThe Lancet (GBD 2019)essay 09, the number that does not move

The years, side by side

YearMPV, all killingsPost, shootings onlyGuardian, all killings
20131,075··
20141,039··
20151,0999951,146
20161,0659591,093
20171,091984·
20181,138992·
20191,099993·
20201,1611,021·
20211,1481,050·
20221,2041,097·
20231,2481,164·
20241,2701,175·

Aggregated from each project's primary dataset · a dot means the project was not counting that year · 2024 is the highest year on both living counts

The readings

  1. 01Who Counts the KillingsThe state's own count was too incomplete to publish, so a newspaper, a nonprofit, and a volunteer started counting.develops the counters
  2. 02The Federal Blind SpotThe FBI has the widest definition of all and, for most of its life, could not publish a national total.develops finding 02
  3. 03Five Counters, Five DefinitionsNo two projects report the same total, because no two are counting the same thing.develops finding 01
  4. 04The Definition GapChange one rule about what counts, and about a hundred people a year appear or vanish.develops finding 01
  5. 05The Open DataA series about counting owes you a count you can open, test, and argue with yourself.develops the release
  6. 06How Many Did Nobody CountMatch the databases person by person and the surprise is agreement. The real undercount turns out to be the government's own death records.develops findings 03-04
  7. 07What a Real Count Would TakeThe design spec assembled from the whole series: a mandate, a written definition, the method the government already validated, and an outside audit. The engineering is done.develops findings 01-06
  8. 08The Two WhysWhy is the count rising, and why do agencies not report? Explanations ranked by evidence strength, ending where they meet: the killings are growing in exactly the agencies that do not report them.develops finding 08
  9. 09The Number That Does Not MoveBlack Americans are killed at two to three and a half times the white rate under every instrument in this series. The measurement, the denominator debate named honestly, and the counting failure that compounds it.develops finding 09
  10. 10Where the Ratio Comes FromThe causes question, graded by evidence: exposure that is demonstrably not race-blind, an encounter stage carried as the live dispute it is, the 1-in-1,000 lifetime risk, and the city-by-city variance that points at policy, not people.develops finding 09
  11. 11The Other Side of the LedgerThe count of what happens to officers fails the same ways: a voluntary assault count computed on whoever shows up, a reporting crater, and volunteer memorials that outrun the official ledger and disagree on definition, 73 versus 458 for the same year. The legibility gap does not take sides.develops finding 02
  12. 12The Word MandatedA career officer said every department is mandated to report deadly force. He was right, and the national count is still voluntary, and both are true at once: department policy, the state patchwork with two working proofs, and a federal mandate the government's own auditors found leaking at every joint.develops finding 02

Live instruments

The ReconciliationThe field-data hub. Five counters side by side, the 2015 definition gap drawn live, and the hashed release with a recompute-it-yourself verify panel.The OverlapThe record-level layer. The three civilian lists nested from MPV's own linkage, the definitional residue decomposed, and why capture-recapture breaks here while the government's vital statistics missed 55.5 percent.The Federal Blind SpotThe one count the government runs, drawn as a coverage curve against the 80 percent threshold it needs before it can publish a total, cleared only in 2024.The Open DataTen tables as CSV and JSON, hashed and versioned, CC-BY-4.0, with the record-level overlap, the verified-claims layer, and the disparity table. Derived aggregates only, never a re-hosted roster of the people who died.
No. 15 · Field DataThe Legibility GapThe empirical companion to the Wartime Edition. Ukraine has assessed roughly 90 billion dollars in housing reconstruction need and disbur...No. 01 · Position PaperClassification as Infrastructure in LLM SystemsEvery planning layer is a classification system. The question is whether we build it deliberately or let it emerge accidentally.No. 12 · Interactive EssayWhat the State KeepsThe American federal state as a portfolio of 49 unexercised options, evaluated through an acquisition optimizer's lens. What gets cut, ke...No. 10 · Interactive EssaySeeing Like an AI CompanySix dead scholars read OpenAI's industrial policy proposal and find six connected problems it refuses to name.No. 17 · Field DataPingree Grove StudiesA civic field-data series about one village, the owner's hometown. Built to be useful to residents first, then to staff and elected offic...

The evidence is 5 independent public counters, reconciled as content-as-code in src/lib/police-shootings/dataset.mjs and frozen into a hashed release (v0.4.0, CC-BY-4.0). Derived aggregates only, never a re-hosted roster of the people who died. Every figure carries a source that resolves: the full provenance ledger ships in the release's download bundle, as machine-readable JSON, and in the claims ledger, and the release re-hashes byte-for-byte in continuous integration.

The twelve essays are dated readings, each frozen to the release current when it was written. The charts above and the instruments are live, re-verify on load, and read the current release (v0.4.0). Data updated 2026-07-23, next review 2026-10. This is a systems-and-design reading of a public-data problem, and takes no partisan side.

the reconciliation hub →/← back to jakelawrence.xyz