How I did it

How I studied Seattle Terry stops

I wanted to know whether parts of Seattle with different racial makeups saw different numbers of Terry stops, once you account for recorded crime, calls to police, how many people live there, and neighborhood conditions. This page covers the data I used, how I put it together, and what it can and can't tell you.

← Back to the story

The data

Six public sources, all from 2020 through 2025 unless I say otherwise.

  • Terry stop records

    Seattle Police Department · 2020–2025

    One record for each stop, including whether the person was frisked or arrested. Some stops show up on several rows, so I counted each stop once. This is the main thing I was measuring: where stops happened and what happened next.

    Keep in mind: The race in these records is what the officer wrote down, not how the person identifies.

  • Calls to police

    Seattle Police Department · 2020–2025

    One record for each call. The public file repeats a call once for every unit that responded, so 3.4 million rows turned into 1.9 million calls. I kept calls from the public separate from calls officers started themselves. Calls from the public give a rough sense of how often people in an area ask police for help.

    Keep in mind: Calls reflect who picks up the phone and why. They aren't a direct measure of harm.

  • Recorded offenses

    Seattle Police Department · 2020–2025

    One record for each offense police recorded and finalized. So I could compare areas with similar amounts of recorded crime.

    Keep in mind: This is crime police recorded, not all crime.

  • Patrol-area map

    Seattle Police Department · Boundaries used since 2019

    The 55 official patrol areas (beats). 51 have residents; the other four cover the harbor and water. Stops are recorded by patrol area, so everything else had to fit onto this map.

  • Who lives there

    U.S. Census Bureau · 2020 Census; 2020–2024 American Community Survey

    Census block counts of residents by race and ethnicity, plus survey estimates of poverty, renting, jobs and age, added up to each patrol area. To describe each area's residents and compare areas with similar conditions.

    Keep in mind: These count residents only, not the commuters, shoppers, students and visitors who fill some areas, especially downtown. Survey figures are estimates, and I carried their margins of error through.

  • Zoning

    City of Seattle · Current zoning

    The land-use zone for every part of the city. To tell mostly residential areas apart from busy downtown and commercial ones.

    Keep in mind: Zoning is a rough stand-in for how many people are around during the day.

How I put it together

  1. Cleaned the stop records

    I counted each stop once, used Seattle's time zone so no stop slipped into the wrong year, and kept the 24,897 stops (94%) that had a standard patrol-area code.

    Cleaning details

    The unit is the distinct Terry Stop ID, because one stop can occupy several weapon rows. Study-period boundaries use America/Los_Angeles; converting from UTC had admitted one late-2019 record and excluded five late-2025 records. 24,897 of 26,483 stops (94.01%) carry conventional beat codes, above the 90% coverage threshold set before analysis. Precinct-month was the documented fallback unit.

  2. Put everything on the same map and calendar

    I matched every source to the 51 patrol areas that have residents, month by month from January 2020 to December 2025. That makes 3,672 area-months, including months with no stops at all, which count as much as the busy ones.

    Seattle has 55 official patrol areas. The other four (99 and H1–H3) cover the harbor and water, so they're in the downloadable map file but not on the public maps.

    Matching details

    • Census 2020 P.L. 94-171 blocks are assigned whole to beats by internal point; largest-overlap and area-weighted allocation are sensitivity checks.
    • Race and ethnicity are mutually exclusive: Hispanic or Latino of any race, plus non-Hispanic single-race and multiracial categories, so shares sum to the total.
    • 2020–2024 ACS block-group counts are allocated by each beat's share of the block group's 2020 population, with 90% margins of error propagated through sums and ratios.
    • Calls are distinct CAD events by Seattle-local month queued (3,436,427 rows become 1,945,173 events). Community-generated and officer-generated calls stay separate.
    • Offenses are distinct offense IDs by offense date. One-month lags and prior-12-month sums are built for violent, property, and all recorded offenses and for community calls.
  3. Compared areas in more than one way

    Raw counts show where stops cluster, but busy areas will always have more. So the maps also show stops per 10,000 residents and stops per 1,000 calls from the public. For what happened after a stop, they show how often stops led to a frisk or an arrest.

    Formulas and map rules

    ComparisonFormulaWhat it shows
    Recorded stopsDistinct stops ÷ years in periodWhere stop records were concentrated.
    Resident rateAnnualized stops ÷ 2020 residents × 10,000Stops relative to the people who live in the area.
    Call rateStops ÷ distinct community-generated CAD events × 1,000Stops relative to one recorded measure of public demand.
    Frisk rateRecorded frisks ÷ stops with known frisk statusWhat happened after a stop was recorded.
    Arrest rateStops resolved as arrest ÷ stops with known resolutionWhat happened after a stop was recorded.

    Pooled 2020–2025 stop counts are divided by six before calculating the resident rate. Call rates use pooled stops and pooled calls. Annual views use that year's stops and calls with the same fixed 2020 resident population.

    The guided map transitions use percentile rank within each comparison so movement is comparable. The explorer uses actual values with fixed six-year color scales capped at the 95th percentile; values above the cap stay visible as numbers and use the brightest color.

    The explorer separates stops coded on-view (an officer saw the event) from stops recorded after a 911, telephone, alarm, text, or later call. Records without a usable call type stay in "All recorded" and aren't assigned to either group.

The models

To see whether racial makeup still mattered after accounting for other differences between areas, I fit five models. Each one makes a different reasonable choice about what to account for. Each number compares otherwise similar areas whose shares of Black residents differ by 10 percentage points. A 1 means no difference; 1.55 means 55% higher. The first four models compare stops per resident; the last compares total stops.

ModelRatio (1 = no difference)95% confidence interval
Race and resident population1.050.62–1.78
Add recorded crime and public calls0.980.58–1.67
Add measured neighborhood conditions1.551.12–2.14
Use the prior year of recorded activity1.20.83–1.73
Model total stops directly1.020.84–1.24

Only the model that includes all the neighborhood conditions gives an answer clearly above 1: about 1.55, or a 55% higher stop rate. The other four ranges all include 1, so they can't rule out no difference. These are patterns across places. They don't show that race caused any stop, and they don't explain why any stop happened.

Model details

All five are negative-binomial models (beat-month stop variance is 9.3 times its mean) with month and precinct fixed effects and standard errors clustered by patrol area. The table reports incidence-rate ratios for a 10-percentage-point contrast in Black resident share.

The full measured-context model includes violent and property offenses, community-generated calls, poverty, renter occupancy, unemployment, age composition, residential density, nonresidential activity zoning, downtown status, Black resident share, month, and precinct. Because race and ethnicity shares are compositional and related to other neighborhood conditions, that coefficient is not interpreted as an effect of race. The large change after removing the resident-population offset is treated as a finding about measurement, not a nuisance.

A separate stop-level diagnostic compares stops of people whom officers classified as Black or White. It excludes every other recorded category, uses the officer's classification rather than self-identified race, and conditions on a stop already having happened. It stays here for transparency, but it isn't featured in the story and can't estimate anyone's chance of being stopped.

Small numbers

When fewer than 20 stops go into a frisk or arrest rate for an area and time period, I don't show the rate. The map says "Not enough recorded stops to estimate" and uses gray, never zero.

  • 14 area-period frisk rates are hidden for this reason.
  • 14 area-period arrest rates are hidden for this reason.

I also left out a year-by-year map filter by the race officers recorded. For 257 of 611 combinations of patrol area, year, and recorded race (Black or White), there were fewer than 20 stops. Those police categories also don't line up with Census categories, so the story never uses who lives in an area as a stand-in for who was stopped.

Confidence intervals

Wilson 95% intervals are calculated before export for every displayed frisk and arrest estimate.

What it can't tell you

It's about places, not people

  • The analysis compares areas, so it can't tell you any one person's risk of being stopped.
  • The records can't show what an officer intended or whether any stop was lawful.

Two different kinds of race data

  • The race in stop records is what the officer perceived, not how the person identifies.
  • Census categories describe residents; police categories describe people who were stopped. They aren't interchangeable.
  • Racial shares add up to a whole, so focusing on one share can hide the rest of a neighborhood.

Rough measures of activity

  • Calls and recorded offenses come out of reporting and policing systems. Neither measures harm without error.
  • Resident counts leave out commuters, customers, students, tourists, and others who are present.
  • Zoning hints at daytime activity but doesn't count the people there.
  • Census and survey context is fixed in time, so it can't show what came first or prove cause.

Downloads and full tables

These files contain only totals by area and time period, plus the official patrol-area boundaries. They contain no stop IDs, case numbers, officer IDs, details about the people stopped, or exact locations.

Complete pooled map-comparison table
Patrol areaAnnualized stopsResident rateCall rateResidentsCalls
B158.6730.569.9519,19835,379
B272.1734.9812.9320,63033,477
B340.3317.528.3623,02728,944
C156.3340.1514.0114,03224,132
C227.1716.328.916,64718,319
C340.3326.1111.5815,44920,900
D1122.17103.321.8311,82633,571
D2119.3369.2921.2417,22333,706
D363.6742.2212.3415,08030,950
E112484.5718.5114,66340,196
E2120.67171.1626.97,05026,914
E387.574.2813.811,77938,053
F174.555.3519.313,46123,162
F290.8392.0622.529,86724,205
F364.3362.2817.8810,33021,588
G157.1778.4414.677,28823,374
G263.1755.314.7211,42325,754
G359.8370.7218.418,46119,504
J130.1719.729.6415,30118,773
J220.678.236.6625,12518,626
J337.3322.4610.216,62021,965
K1103.5459.823.442,25126,498
K2118470.4925.942,50827,298
K3483.17843.6650.995,72756,850
L145.6724.2710.1818,81626,928
L256.1730.8611.3718,20029,646
L35129.7911.3417,12026,981
M1114.17208.7527.585,46924,834
M279.17228.8125.783,46018,428
M3303.833,360.9970.3790425,907
N12211.367.0119,36118,843
N2104.33103.817.610,05135,576
N310881.5618.1813,24235,647
O160.33977.852161717,237
O247.831,339.8717.2635716,627
O364.83246.722.472,62817,313
Q163.8327.0220.3223,62418,846
Q272.17251628,87227,062
Q3128.3388.1321.0314,56136,606
R159.1731.912.3718,54528,693
R2124.1785.7717.6514,47642,215
R353.3330.9212.9217,25124,775
S144.6726.8212.2816,65721,816
S251.8339.6313.4913,07823,061
S345.6733.951313,45221,069
U157.6721.378.4926,98540,775
U264.6731.411.120,59234,947
U332.838.716.1237,69132,208
W152.3327.711.4818,89027,353
W268.6727.5813.5924,89430,313
W337.831712.4222,25618,279
Outcomes by racial classification entered by officers

These administrative categories reflect an officer’s perception, not self-identification. Rates describe outcomes among recorded stops and cannot estimate a person’s chance of being stopped.

Recorded categoryStopsFrisk rateArrest rate
American Indian or Alaska Native54626.2%28.8%
Asian1,01028.9%23.7%
Black7,62731.6%26.9%
Native Hawaiian / Pacific Islander16926.6%24.9%
White12,11224.3%22.8%
Unknown2,90030.2%17.9%
Not recorded53224.6%6.2%
Multiple subjects1Not enough recorded stops to estimateNot enough recorded stops to estimate
Complete resident race and ethnicity composition by patrol area

“Hispanic or Latino” is reported separately. The remaining Census categories in this table are non-Hispanic.

Patrol areaHispanic or LatinoWhiteBlackAmerican Indian / Alaska NativeAsianNative Hawaiian / Pacific IslanderAnother raceMultiracial
B16.9%71.2%1.9%0.4%11.9%0.2%0.6%6.9%
B26.1%74.6%1.9%0.4%8.7%0.1%0.6%7.6%
B36.5%72.3%1.8%0.3%10.3%0.1%0.6%8.1%
C17.4%71.8%2.7%0.2%9.9%0.2%0.6%7.2%
C26.2%73.3%4.1%0.3%8.5%0.1%0.6%7%
C36.4%66%9.1%0.2%9.3%0.2%0.6%8.2%
D17.4%50.1%5.4%0.5%30.5%0.2%0.7%5.2%
D27.3%43.6%5.1%0.3%37.4%0.3%0.7%5.3%
D36.9%54.3%3.9%0.4%28.3%0.1%0.6%5.6%
E19.4%63%4.4%0.3%14.4%0.2%0.8%7.6%
E29.6%56.1%6.8%0.6%18.3%0.5%0.6%7.5%
E38.9%58.1%5.7%0.7%19.5%0.3%0.6%6.2%
F110.2%45.4%17.4%0.8%16.6%0.6%0.5%8.4%
F217.6%49.5%10.2%0.6%12.7%0.7%0.7%8%
F320.8%44.5%9.1%1.1%15.6%1%0.5%7.5%
G110.4%42.2%14.1%0.7%24.4%0.5%0.6%7.1%
G29.5%51.5%17%0.5%12.4%0.3%0.7%8.2%
G38.1%50.3%15%0.4%16.4%0.3%0.5%9.1%
J17.2%69.7%3.1%0.6%10.4%0.3%0.8%8%
J25.5%79.2%0.8%0.3%6.2%0.1%0.5%7.5%
J36.5%73.3%1.3%0.3%9.9%0.1%0.6%8%
K18%55.1%21.3%1.4%9.1%0%0.4%4.7%
K211.1%57%13.6%2.3%9.2%0.7%0.2%5.9%
K37.7%29%12.5%1%43.1%0.2%0.6%5.9%
L111.2%52.5%10.6%0.5%16.5%0.3%0.6%7.9%
L28%62.7%5%0.2%15.2%0.3%0.8%7.8%
L37.6%62.4%6.4%0.5%14.8%0.1%0.7%7.4%
M17.6%59%6.1%1.1%20.3%0.3%0.7%5%
M27%52.1%5%0.4%29.7%0.1%0.7%5%
M38.6%55.9%5.8%1.2%20%0.9%0.7%7%
N16.8%70%5%0.3%9.7%0.2%0.9%7%
N212.4%50.7%12.1%0.7%15%0.3%0.7%8.1%
N312.3%55.3%6.5%1%16%0.3%0.5%8.1%
O118.8%43.8%10.4%6.8%5.2%2.1%1.9%11%
O219.3%50.1%9%0.6%7%0%1.7%12.3%
O311.1%46.1%6.3%0.7%27.7%0.5%0.6%7%
Q17.2%72.8%1.7%0.3%9.8%0.3%0.6%7.4%
Q27%71.8%2.1%0.4%10.5%0.2%0.5%7.5%
Q38.4%60.2%4.2%0.4%20.2%0.1%0.7%5.8%
R114%37.1%6.9%0.5%32.9%0.3%0.4%7.9%
R27.4%39.6%24.3%0.5%20.1%0.3%0.5%7.3%
R38.1%52.5%13.6%0.3%16.1%0.3%0.9%8.1%
S16.5%15.3%22.5%0.2%49.7%0.5%0.4%4.8%
S211.2%22.3%24.4%0.3%34.8%0.6%0.6%5.8%
S310%25.7%27.1%0.4%29%0.7%0.4%6.8%
U17.9%63.6%2.6%0.3%17.3%0.1%0.6%7.6%
U28.5%38.2%3.6%0.5%41.6%0.4%0.4%6.7%
U35.4%69.6%2.2%0.3%13.6%0.2%0.7%8%
W16.7%76.2%2.4%0.4%6.6%0.1%0.5%7.2%
W27.6%72.2%3.3%0.3%7.9%0.3%0.5%7.9%
W37.4%73.7%4.4%0.5%5.7%0.3%0.7%7.3%