New York City gunfire, 2006

The trend

Down, then up, then down again

Shooting incidents per year. The 2010s decline gave way to a pandemic surge in 2020–21 — both full, comparable years — and then the steepest fall on record. 2026 is partial (through March).

Y-axis

The toll of the decline

Had 2006 levels held, thousands more would have been shot

A counting exercise: take 2006's tally — people shot — and hold it flat through 2025. The gap between that line and what happened is how many fewer people were shot over two decades of decline, net of the pandemic years, when the real line rose back above 2006.

Fewer people shot, 2007–2025

Gun violence within serious violence

A shrinking share of serious violence involves a gun

Shootings are not a standalone crime category. In the New York Police Department's seven "major felonies," a shooting is usually folded into one of two buckets: a victim who survives is generally recorded as a felony assault, one who dies as a murder. So shootings are best read as the gun-injury slice of serious violence. That is why the two are charted at their own, very different scales below (felony assault is an order of magnitude larger) rather than on a shared line. Compare the two: for every 100 felony assaults in 2006 there were about shootings; by 2025, about .

Felony-assault counts are all felony assaults citywide (NYPD complaint data, dated by report date), the great majority of which do not involve a gun. The "shootings per 100 felony assaults" ratio is a rough index of how gun-driven serious violence is — not a clean subset share, since the ~16% of shootings that are fatal are counted as murders, not assaults.

What changed underneath

The totals fell, and the patterns shifted

Three things changed over the twenty years:

Every neighborhood, every year

The whole city on one grid

One row per neighborhood, one column per year, each cell shaded by its shootings. Grouped by borough. Read across a row to follow one place through two decades; read down a column to see a single year citywide. Switch to per 100,000 residents to rank by intensity rather than sheer numbers — which lifts the Bronx and Brownsville to the top.

Shade by

Sort

Where it concentrates

The geography of risk, today

This section shows the current level of shootings. Raw counts and per-capita rates disagree: the busiest precinct by sheer numbers is large; adjusted for its population, the highest rates are elsewhere.

The shifting geography

Which neighborhoods changed the most

Now change, measured three ways; toggle between them. Precincts labeled by the neighborhood they cover.

Compare →

The map

The map, year by year — or all at once

Drag the slider to move through the years, or press play to watch the decline unfold. Or switch to 20-yr change to see the whole trajectory on a single map: where shootings fell most (steel) and the few places they rose (red), 2023–25 versus 2006–10.

Show

Measure

Fatal only

All years

Streetlights & gunfire

Do broken streetlights track the shootings?

A common intuition — dark blocks breed crime — predicts two things: that the worst shooting blocks should be where streetlights stay broken, and that shootings should cluster when the lights are out. Both are tested on a matched window — shootings and "street light out" 311 complaints over the same 2010–2026 period, on a ~110-metre block grid. (311 streetlight complaints exist back to 2004, but the pre-2010 files carry no coordinates — only ZIP and borough — so they can't be placed on a block; geocoded records begin in 2010, which sets the window.) Both tests give the same answer: the lights and the gunfire don't track each other. Year by year, a block saw about the same number of shootings in years its light was reported out as in years it wasn't (a ratio of ); of shootings fell on a block with a same-year outage, against a baseline. And across blocks the overall relationship is flat (correlation ) — only of the worst shooting blocks are also chronic-outage blocks.

Each dot is a high-shooting block · outage complaints vs shootings

The blocks that are both

the minority of hotspots that are also chronic-outage blocks

A caveat that cuts the other way: 311 counts reported outages, and reporting is uneven — lower-income, higher-violence neighborhoods tend to file fewer 311 complaints per actual problem. So a low outage count on a shooting block may reflect under-reporting, not working lights. This compares where outages are logged, not necessarily where lights are dark. Both layers also rise with population. Treat it as descriptive, not causal.

The people

Who was shot, and who fired

Victim records covered essentially every shooting (28,912). Offender records existed only where a suspect was identified (18,901) — an arrest-skewed subset, read with care. Shown side by side.

Victims

28,912 people shot · age, sex, race

Offenders

18,901 identified suspects · age, sex, race

How many at once

Most shootings hit one person; some hit several

The data records no count of shots fired, so the closest measure of an incident's severity is how many people it struck. Most shootings have a single victim; a minority wound two or more.

People shot per incident · all incidents, 2006–2026 · "5+" caps the tail

The clock and the calendar

When shootings happen

Summer, weekends and late nights — though, as the cards above showed, the late-night concentration has eased over the years.

By day and hour

share of all shootings in each weekday-hour cell — darker is more

By month

seasonality, all years

By day of week

Monday through Sunday

Methodology, sources & limitations

Sources. NYC Open Data's shooting collection (published Feb 2026): Shootings (5ucz-vwe8), Victims (pztn-9bne), Offenders (gdk4-mbsv), and the shoreline-clipped Police Precincts (y76i-bdw7). Records run ; 2026 is partial.

Counting. "Shootings" are incidents (one row per event); victims/offenders join on incident_key. An incident is fatal if any victim carries the statistical-murder flag. The source latitude/longitude columns are swapped; we correct them before mapping. Junk ages (e.g. 1022) and mixed sex labels are normalized.

Neighborhoods & change windows. The geographic unit is the NYPD precinct, labeled with the neighborhood it mainly covers (approximate, for readability). Share of city total compares a pre-pandemic window (2015–2019 average) with a recent one (2021–2025). Pandemic spikes compare 2018–2019 with 2020–2021. Declines compare 2020–2021 with 2023–2025. The share comparison is also user-selectable to any two years (2006 onward). Low-volume precincts (under 60 lifetime shootings) are excluded from the change tables for stability.

Population & rates. Per-capita rates use actual precinct population in each year, not a fixed figure. We took total population by census tract from the 2010 and 2020 decennial censuses, assigned each tract to a precinct by its centroid, summed, and interpolated between (and extrapolated beyond) the two censuses. This captures real local change — the Financial District precinct grew ~130% and Downtown Brooklyn ~70% between censuses. The all-precinct 2020 total reconciles to ~8.77M. Central Park's precinct is omitted from rates (almost no residents).

Confidence. High for incident counts, geography, timing and victim demographics; medium for offender demographics (identified-suspect subset) and the demographic-share trends (category definitions can shift over twenty years). Nothing is modeled — every figure is a direct count of published records.

How this was made. The pipeline, maps and charts were built with AI assistance (Claude). Every number is computed directly from the four datasets by a script; nothing is AI-estimated. Checks: incident/victim/offender counts reconciled with source totals (24,127 / 28,912 / 18,901); the lat/lon swap caught and corrected against sample records and the precinct map; yearly totals and the fatal-incident join spot-checked; precinct change windows hand-verified for several neighborhoods. Not independently re-reported: NYPD classification, geocoding and arrest decisions, taken as published.