How this map was built

This document records every data source, every analytical choice and every known limitation behind the map of New York City misdemeanors and violations. The two scripts that produce the data are public and re-runnable.

What the map shows

Every record New York City's Police Department classifies as a misdemeanor or a violation — the two lowest tiers of offense, below felonies — from January 1, 2015 through March 31, 2026, aggregated to the city's 78 police precincts. Felonies are excluded entirely. Two separate lenses are offered:

• Complaints — incidents reported to or recorded by police (the standard "crime" dataset). Measures reported crime.
• Arrests — arrests made by police. Measures enforcement activity. An arrest is itself an enforcement action, so this lens responds directly to "enforcement pushes."

Data sources

All data comes from New York City Open Data and was retrieved on May 21, 2026.

Dates

Complaints are dated by report date (RPT_DT), the date the incident was reported to police — not the date the offense is alleged to have occurred (CMPLT_FR_DT), which can be years earlier for some offenses. Arrests are dated by arrest date (ARREST_DATE).

The enforcement-sensitivity classification

The map separates offenses whose recorded counts move with policing strategy from those that mostly reflect underlying victimization. Each offense type is assigned to one of three groups. This is a documented editorial judgment, not an official NYPD category — the complete assignment is exported to classification.csv so it can be inspected or contested.

Enforcement-sensitive Discretionary, officer-initiated offenses. A police officer largely decides whether these become a recorded incident, so the counts rise and fall with proactive enforcement: dangerous drugs, cannabis offenses, criminal trespass, theft of services (fare evasion), gambling, loitering, prostitution, disorderly conduct, weapon possession, intoxicated/impaired driving, vehicle-and-traffic-law offenses, alcoholic-beverage offenses, fraudulent accosting, jostling, burglar's tools, fortune telling.

Complaint-driven Offenses typically reported by a victim or witness, where a member of the public, not an officer, usually initiates the record: petit larceny, assault 3, harassment, criminal mischief, sex crimes, forcible touching, fraud, unauthorized use of a vehicle, possession of stolen property, offenses against children, endangering welfare.

Other / mixed Administrative catch-alls, ambiguous grab-bag categories and offenses that arise downstream of police contact rather than from a push to police a behavior: offenses against public administration (resisting arrest, obstructing governmental administration), "offense against public order/sensibility," administrative code, miscellaneous penal law, other state laws, anticipatory offenses, escape, agriculture & markets law, health code, uncoded entries.

Why some judgment calls went the way they did

• "Offenses against public administration" (resisting arrest, obstruction) is placed in Other, not Enforcement-sensitive. Its counts are driven by police contact, but it is a consequence of enforcement rather than a target of an enforcement push, so grouping it with drugs or trespass would overstate the enforcement-sensitive total.
• "Offense against public order/sensibility" is a large, heterogeneous bucket (public lewdness, unlawful surveillance and more) that is part proactive and part victim-reported. It is left in Other.
• Dangerous weapons (possession) and vehicle-and-traffic-law offenses are treated as enforcement-sensitive because, at the misdemeanor/violation level, they are overwhelmingly discovered through stops and proactive checks.

Main limitation: summonses are missing

More generally: complaint data measures reported crime, which is shaped by public willingness to report; arrest data measures police activity, not the amount of crime occurring. Neither is a direct measure of how much low-level offending happens.

The complaint-to-arrest ratio view

The map's third shading option divides each precinct's complaints by its arrests (for the selected offenses and offense level, in the selected year, using both lenses at once). A high ratio means reporting outpaces enforcement; a low ratio means the precinct is enforcement-heavy relative to how much is reported.

Because complaints structurally outnumber arrests almost everywhere, the diverging color scale is centered on the citywide ratio for that year and selection (about 2.4 complaints per arrest citywide in 2025), not on 1.0. Red shades a precinct that is more arrest-heavy than the city as a whole; blue shades one that is more complaint-heavy. The scale spans roughly 1.8× either side of the city average on a log scale. Precincts with zero arrests for the selection are treated as maximally complaint-heavy; precincts with no records are left uncolored.

Geographic method

• Incidents are aggregated to the precinct of occurrence (ADDR_PCT_CD for complaints, ARREST_PRECINCT for arrests). Records with no precinct are dropped.
• Precinct 116 (operational since 2024 in southeast Queens) is shown on its own polygon. Precinct code 483 appears on three stray records and is dropped as invalid.
• The default count view is raw incidents, not adjusted for population, so a precinct with more people, more commercial activity or more transit tends to show higher counts. The rate per 1,000 residents view (below) adjusts for population; the enforcement-sensitive share view is a within-precinct ratio and so is also insensitive to raw volume.

Per-capita rates (population that varies by year)

The rate per 1,000 residents shading divides a precinct's incidents (for the selected lens, offense level, offenses and year) by its residential population, then multiplies by 1,000. Because NYPD precincts are not census geographies and precinct populations change over time, the denominator is built to be both real and time-varying:

• Source. Residential population comes from the 2010 and 2020 decennial censuses, apportioned to 2020 NYPD precinct boundaries by John Keefe's census-by-precincts crosswalk. Both censuses use the same 2020 boundaries, so the two counts are directly comparable.
• Between the censuses (2015–2020). Each precinct's population is linearly interpolated between its 2010 and 2020 counts, giving a distinct estimate for every year, anchored to the two census counts.
• After 2020 (2021 onward). Population is held at the 2020 count. No reliable annual population exists at the precinct level after the 2020 census, and extrapolating a trend can overshoot, so the map holds the last census count. Recent rates should be read with that caveat: the citywide population moved modestly after 2020, but the map keeps precinct denominators fixed.
• Excluded precincts. The 22nd (Central Park) has a negligible residential base (about 130 people) so a per-capita rate is meaningless; the 116th (operational since 2024, carved from the 105th) has no census crosswalk yet. Both are shown grey in rate mode and left blank in the CSV.
• Daytime-population flag. Midtown precincts 14 and 18 run high on per-capita rates because their daytime and commuter population dwarfs the residents the census counts. The tooltip and dossier flag this; the rate is real but reflects a denominator that undercounts who is present.

Rates are a comparison aid, not a risk score. The full per-precinct, per-year population and rate for any selection is available through Download CSV.

Data cleaning

• Offense names are canonicalized: the source data contains truncated and variant spellings of the same offense (for example CRIMINAL MISCHIEF & RELATED OF and CRIMINAL MISCHIEF & RELATED OFFENSES), which are merged to one label.
• Offense level is read from LAW_CAT_CD (MISDEMEANOR/VIOLATION in complaints; M/V in arrests). Felony, infraction and unclassified levels are excluded.
• All aggregation is done server-side through the Socrata API (GROUP BY queries); no raw person-level records are stored or shipped.

Reproduce it

Two scripts regenerate everything:

• fetch_data.py — pulls precinct × year × offense × level counts from the four NYPD datasets. It auto-detects the latest complete calendar quarter from the year-to-date datasets (so nothing has to be hand-edited when a new quarter lands) and writes the resolved window to data/meta.json.
• build_data.py — canonicalizes offenses, applies the classification, attaches per-year populations and neighborhood names, simplifies the precinct boundaries and writes data.json, precincts.json and classification.csv.

Run python3 fetch_data.py && python3 build_data.py. No API key is required. A monthly GitHub Action re-runs both and commits the result only when the data changes.