Questions about New York, answered from public data
The city publishes an enormous amount about itself, in hundreds of tables that
were never meant to be read together. Most of what gets
written from them is an average and a narrative, and the average hides most of
what is going on. Each article here takes one question, finds the datasets that
bear on it, and works out what they can and cannot say — with interactive charts
you can cut your own way, the data, and the code that produced every number.
All models are wrong; some are useful. We spend our time on the second clause.
In January to May 2026, felony assault per subway entry was about twice its 2019 rate, counting unpaid entries, while the six major felonies together sit close to 2019.
Is the subway safer? Early this month the NYPD called 2026 so far the second safest year in its transit system in fifteen, the pandemic years left out. Days later two stabbings put riders on alert. Both can be true. Per ride, the police record a third as many major felonies as in 1998, and the year to June sits close to 2019 once fare evaders are counted as riders. One crime went the other way. Felony assault runs at about twice its 2019 rate, and it was climbing before the pandemic emptied the trains. Two things often said about the trains, that crime is only down because fewer people ride and that a wider law made the assaults, get tested against the counts. What they can't say is the question riders ask: why did assault go the other way?
New York's thirty-year fall in road deaths belongs to people on foot and in cars. Deaths on two wheels rose, to about 71 a year in the early 2020s.
New York counted 205 traffic deaths in 2025, the fewest since its count began in 1910. Over thirty years of federal records, deaths on foot and in cars fell by half or more. Deaths on two wheels did not: one in ten of the dead in 1995, nearly three in ten in 2024. The police records put the new deaths on e-bikes, e-scooters and mopeds, a vehicle the Vision Zero plan of 2014 doesn't name. We set the counts beside the growth of app delivery, the legalisation of 2020 and the new 15 mph limit. Then we test what was said at the Council's hearing, starting with the claim that e-bike deaths are up again this year. The question the crash records can't answer on their own is the one the city most needs answered: has riding got riskier, or are there simply far more riders?
Shootings in 2025 were the fewest in the NYPD's open files, which begin in 2006; recorded felony assaults passed 1998's count, on a definition that has widened.
New York's fall in gun violence is real, but it isn't the whole story. Shootings in 2025 were the fewest in the NYPD's open files, and murders this year are running a fifth below last year's. Yet felony assault, the most common violent felony, is up 44% since 2019 and past the NYPD's own 1998 benchmark. Why are the two moving apart? And why, while shootings fall, were nearly twice as many people under eighteen shot as in 2019? We follow both series through thirty-five years of records and ask how much of the rise the law's wider definition can explain. Then we test four things New Yorkers say about the violence, starting with the bail law. The one thing the files can't say is the one everyone argues about: why.
Last winter's record of about 305,000 no-heat complaints was not just the cold: at the pre-pandemic rate its weather would have brought about 206,000; it brought 48% more.
New York's heat season opened on 1 October after a second record winter in a row: about 305,000 complaints of no heat, in the coldest winter in eleven years. Its weather alone would have predicted about 206,000. Inspectors found more than twice as many heat violations as in a winter before the pandemic, and more of the complaints come from the same buildings each year. And the record barely sees public housing. There, the City's own survey says, nearly four in ten households had their heat break down for six hours or more in the winter of 2021-22; almost none of it reaches HPD's complaints. Thirty winters of records say who stays warm in New York. Whose cold goes uncounted?
Ambulance response to calls ultimately classed as life-threatening rose by a minute and three-quarters in fiscal 2026, to the slowest year in the city's figures since 1988. The extra time appears before and after an ambulance is assigned, not in the 911 call. For calls urgent from the start the rise was about a minute, most of it after assignment, as ambulance capacity fell and calls did not rise.
The Mayor's Management Report says ambulances took 10:32 on average to reach life-threatening medical emergencies in fiscal 2026, up 1:43; the FDNY's dispatch records reproduce it within five seconds (10:37, up 1:45). In the city's own figures, which run back to fiscal 1988, it is the largest one-year rise and the slowest year since 1988: thirty years ago, in fiscal 1996, the year the Emergency Medical Service merged into the Fire Department, the figure was 8:58, and it fell to a low of 6:25 in fiscal 2012 before rising five years running. Call-taking barely moved (five seconds); the time from job creation to first assignment rose 51 seconds and the trip from assignment to arrival 54. That headline counts calls by their final coding and times upgraded calls, about one in eight, from the original call. For calls life-threatening from the start the rise was 56 seconds (57 under the initial-coding definition), with under ten seconds before assignment and 47 after; upgraded calls account for 64 of the 105 seconds in a shift-share decomposition, and the public file cannot separate their queueing from time spent at a lower priority. Calls held because no unit could be assigned at once rose from 6.9% to 11.5%. Ambulance in-service hours fell 5.4% while all ambulance incidents fell 0.6%, about 5% more calls per unit of capacity; hospital turnaround was flat and fire engines on the same kind of call slowed by eleven seconds, so the evidence converges on tighter ambulance capacity, though the data do not identify the cause. Every borough slowed, and the summer of 2026 was slower still.
A research publication about New York City, built on the public record: the
Census Bureau's surveys and microdata, the city's own housing, shelter and tax
data, the Federal Reserve, the Bureau of Labor Statistics, and whatever else a
question turns out to need. Every article starts from a number someone has
already published and ends with what that number was hiding.
What is published about a city is a collection of slices, each careful and
correct, no two cut from the same block. The interesting questions live in the
gaps between them, and in the intricacies that a citywide average smooths
over. Finding out how far the data can be pushed into those gaps, and saying
plainly where it stops, is the whole job. Rather than offer one narrative, an
article gives a general view and the controls to explore the facets that
interest you.
How an article is built
Start from the published table. Quote the official number,
reproduce it from the underlying microdata, and reconcile the two, measuring
and explaining any difference, before going anywhere the table does not.
Cut the microdata the way the table cannot. Household by
household, with the survey's own margins of error on every bar.
Bring in the second dataset — tax schedules, a housing
survey, a shelter census — and say exactly how it was joined and what that
join assumes.
Separate what is measured from what is inferred. A
relationship observed in one survey is a finding. A relationship stitched
across two is an estimate, reported with a range or not at all.
Ship the working. The dataset, the script that computes
every figure, and the charts you can filter yourself.
Why the interesting questions are hard
The census knows a household's income and rent but not whether its lease is
stabilised. The city's housing survey knows the lease but has a tenth of the
sample. The Federal Reserve knows the balance sheet but only for the nation. No
table crosses all three, so a question like "who really lives in a
rent-stabilised apartment, and what do they own?" has no published answer.
Most of the time the honest answer is partial: this much can be measured,
this much can be estimated under a stated assumption, and the rest is not
knowable from public data. Articles here say which is which, in the text and
on the chart.
What every article promises
Sourced
Every quoted figure links to the table it came from, with the vintage
and, where the source supports one, the margin of error. Nothing is typed
in from memory.
Reproducible
The numbers in the text are computed by the shipped script from the
shipped data. Rerun it on the same inputs and you get the same numbers.
Labelled
Measured, estimated or assumed — each number says which, and every chart
says what it cannot show.
How the work is done
A lot of the research is done with the assistance of a combination of AI
agents: pulling and harmonising datasets, drafting and running the analysis
code, checking figures against their sources. A project of this scope would
not be possible for one person otherwise. What they do not do is decide what
is true.
There is a human in the loop at every stage. Every article is read, checked and
proofread by an actual expert in the field before it is published: a data
scientist with over ten years in the financial industry, anonymous for now to
keep the focus on the models and the procedures. The tools are there to manage
the scope, not to add to the slop. The site starts with New York because it is
a dynamic and unusually well-documented city; it may wander further later.
More about the site.
Swipe or use ← →
SparkyData Weekly
Once a week, one New York theme: what the public data says about it, and the week's new articles.
Free, weekly, and one click to leave. Your address is held by Buttondown, which sends
it; see the privacy policy.
