The disparity survives every definition. This essay asks why it is what it is.
The last essay ended on a discipline: measure first, attribute carefully or not at all. This is the attribution essay, so the rules tighten rather than loosen. Every claim below is quoted from a published study and carries a chip grading how strong its evidence is. Where the literature's most famous result and its most famous rebuttal collide, both are carried, as the live dispute they are. Everything stays at the level of systems, geography, and incentives. No verdict is passed on any officer, department, or policy, and where the research cannot support a conclusion, this essay says so instead of supplying one.
Start with arithmetic, because it organizes everything that follows. A group's rate of being killed by police is the product of two numbers: how often its members encounter police, and how those encounters go once they begin. A 2.5x to 3.5x population-level disparity can come from the first factor, the second, or both, and every study in this literature is an attempt to split the product. Keep that frame and the famous disagreements stop being contradictions; they are mostly measurements of different factors.
The best-documented layer is the first one: Black Americans have far more police contact per capita, and part of that gap is demonstrably not produced by behavior.
The largest study of American police stops ever assembled, roughly 95 million traffic-stop records from 21 state patrols and 35 municipal departments, found Black drivers stopped at higher rates, and then ran the elegant test: compare stops made in daylight, when an officer can see a driver's race, against stops after dark, when a "veil of darkness" hides it. The Black share of stops falls after sunset. The same study found the evidentiary bar for searching Black and Hispanic drivers ran lower than for white drivers.[src] A stop rate that changes with the visibility of race is responding, at least in part, to race.
Where stop programs kept full records, the concentration is not subtle. At the 2011 peak of New York City's stop-and-frisk program, officers recorded 685,724 street stops; about 87 percent of the people stopped were Black or Latino, and roughly nine in ten were neither arrested nor ticketed.[src] Each stop is a small draw from the escalation lottery. Run vastly more draws on one group and the group's death rate rises even if every single draw carried identical risk.
The deeper layer is why contact concentrates where it does. Police deploy heavily where violent crime concentrates, and in America that geography is not a natural fact: a century of explicit federal, state, and local policy, redlining, restrictive covenants, public-housing siting, highway routing, concentrated Black Americans into high-poverty neighborhoods, a history documented at book length in Richard Rothstein's The Color of Law.[src] The consequence for this series is uncomfortable and precise: even a perfectly race-blind "police where the crime is" strategy inherits a racially built map. Part of today's exposure disparity is yesterday's housing policy, laundered through geography.
The second factor, whether encounters themselves go differently by race, is where the literature genuinely fights, and this series' own subject matter, nested lists and broken denominators, turns out to be the whole fight.
The most cited paper on the other side is the economist Roland Fryer's 2019 study, which found that conditional on an encounter, officer-involved shootings in its data, chiefly Houston, showed no racial difference, while sub-lethal force showed a large one: Black and Hispanic civilians were more than 50 percent more likely to experience hands-on force in a police interaction, a gap that survived the study's controls.[src]
The most cited answer, by Knox, Lowe, and Mummolo in 2020, is not a competing dataset but a proof about method: if the stops themselves are racially skewed, then studying only recorded encounters filters the two groups differently before the comparison begins, and the standard approach can severely underestimate encounter-stage bias or mask it entirely.[src] Readers of this series have met this trap before: it is the same class of error as running capture-recapture on nested lists in essay six, a denominator produced by the very process being measured. And the veil-of-darkness result above is direct evidence that the precondition, skewed stops, holds. The dispute is live; a 2025 correction to the rebuttal paper fixed a coding error without changing its conclusions, and researchers continue to argue the bounds.
Two population-level observations sit awkwardly with a pure "encounters explain everything" account. Mapping Police Violence's 2024 analysis found Black Americans killed by police were more likely than white Americans to have been unarmed and not attacking anyone when killed.[src] And the disparity appears in encounter types with no crime nexus at all, the traffic stops above being the largest example. Neither observation settles the decomposition. Both constrain it.
In 2019 the demographers Frank Edwards, Hedwig Lee, and Michael Esposito converted the counts this series tracks into a lifetime risk, and it is the single most clarifying number in the literature: about 1 in 1,000 Black men can expect to be killed by police over the life course, roughly 2.5 times the risk carried by white men, with risk peaking between ages 20 and 35, where police violence ranks among the leading causes of death for young Black men.[src]
The same body of work carries the second clarifying fact: the disparity varies enormously by place. Some metropolitan areas run far above the national ratio, others near parity. That variance is itself evidence about causes. A disparity produced by something constant in people would not swing by a factor of several between cities; a disparity produced by segregation patterns, deployment choices, and department practice would, because those are exactly the things that differ between cities. The ratio moves where policy moves. That is the strongest single clue about where it comes from.
The most common counterargument deserves to be met at full strength. It says: benchmark killings against crime involvement or arrests instead of population, and the disparity shrinks, so the exposure difference is just a crime difference. The answer runs through everything above. Arrests and recorded encounters are not a neutral denominator; they are produced by the stop-and-deployment layer this essay just documented as skewed, which is the Knox-Lowe-Mummolo point in one sentence. The disparity persists in encounters with no crime nexus. And "the exposure is the mechanism" does not dissolve the disparity; it relocates it one step upstream, into decades of documented policy about where Black Americans could live and how those places would be policed. Essay nine named the denominator debate without settling it, and that discipline holds here too: benchmarks answer different questions, and no benchmark makes the population-level burden untrue.
The fairest version of the benchmark deserves its own numbers, and the government publishes them. The Bureau of Justice Statistics surveys police contact directly: in 2022, about 49 million US residents sixteen or older, 19 percent, had at least one contact with police, and 8 percent experienced contact the police initiated.[src] The overwhelming majority of those contacts involve no force of any kind, which is the true and important fact the benchmark argument starts from. But read one more row of the government's own table: among people whose most recent police-initiated or accident-related contact was surveyed in 2022, 2 percent of white respondents reported the threat or use of nonfatal force, and 6 percent of Black respondents did.[src] Three to one. The survey built to supply the encounters denominator reproduces, inside the encounters themselves, the same ratio this essay has been decomposing. Even the counterargument's own instrument reports it.
This series' own contribution to the causes question is the one nobody disputes: whatever produces the disparity, the state's instruments were worst at recording it. Death certificates misclassified 59.4 percent of Black Americans' deaths from police violence, the highest share of any group, against 55.5 percent overall.[src] For forty years the official record understated not just the total but the gap, which means every argument in this essay had to be had on data the public built. A problem the state cannot see is a problem the state cannot answer, and the seeing failed hardest exactly where the burden fell heaviest.
What this essay has said: the disparity decomposes into exposure and encounter; the exposure layer is large, well-documented, and partly race-responsive by direct test; the encounter layer is real for sub-lethal force and genuinely contested for lethal force, on methodological grounds this series' readers already understand; the lifetime burden is about 1 in 1,000 for Black men; and the ratio's variation across places points at policy variables, not people.
What this essay has not said: that any particular officer, department, or policy caused any particular death, that the decomposition is settled, or that intent can be read off a rate. Disparity is a measurement; discrimination is a finding about mechanism; the studies above are the record of researchers trying to get from the first to the second, honestly and in public. This essay carries their state of play, not a verdict.