Not just the cold: thirty winters of heat in New York

From 1 October to 31 May, the heat season, building owners must keep apartments at least 68 degrees Fahrenheit between 6am and 10pm whenever it is below 55 degrees outside, and at least 62 degrees between 10pm and 6am whatever the weather. Hot water must be at least 120 degrees all year round, HPD says.

A tenant without heat can complain through 311, by phone, online or in its app. Since this season began, inspectors must try to see every apartment that complains and gives a name, so it is worth complaining even if a neighbour has. Repeat calls from one apartment are still duplicates while the first complaint is open, and an anonymous complaint is a duplicate once a complaint naming an apartment is in. NYCHA residents report to NYCHA, through the MyNYCHA app or its Customer Contact Center at 718-707-7771.

A Council bill introduced in January 2026, Int 0252, would raise the minimums to 70 degrees by day and 66 at night; it was still in committee at the start of October.

Closely, everywhere we could check. HPD’s complaint file and 311’s own record count the same requests: summed over all heat and hot-water complaints, their season totals agree within 2.5% in 16 of the 22 seasons since 2004-05 and within about 5% in all of them (311’s yearly files misplace some December requests). The City’s report gives over 300,000 heat or hot-water complaints in 2025 and over 225,000 in each year since 2023; the file gives 318,652 for 2025 and at least 230,610 in each of those years. HPD’s heat page counted 344,437 heat complaints in the 2025-26 season to 31 May. That count takes in hot water as well as heat: our count of every heat and hot-water problem over the same months is 343,707, a few hundred fewer, and the no-heat problems we follow throughout are 304,578 of them.

In January 2026, 311 recorded 79,928 heat and hot-water complaints, the “nearly 80,000” Gothamist reported; HPD told Gothamist that about 37,000 remained without duplicates, the most for any month. HPD’s file agrees that it was a record month, with or without hot water: counting no heat alone, it brought 74,168 complaints, ahead of January 2004’s 69,553.

We count HPD’s problem records: a complaint can list more than one problem, and we count each no-heat problem once. Distinct complaints run slightly lower (0.4% lower last winter, and at most about 3%, in 2011-12), and every record here holds on either count.

The survey reproduces too. The microdata give, to the household, the Census Bureau’s published counts of renter households with and without a heating breakdown in every table it published (1999 to 2014), HPD’s 1,719,167 renter households without a breakdown in 2017, and HPD’s 2021 counts of breakdowns among all, rent-stabilized and private unregulated households within their margins.

Partly, and the answer depends less on how the weather is allowed for than on what else is. The benchmark in the first chart assumes complaints rise in step with heating degree days, at the 43.0 per degree day of the sixteen winters from 2003-04 to 2018-19. On it, last winter’s cold would have brought about 206,000 complaints, nearly 100,000 fewer than it did: 48% above. We stop before 2019-20, the winter the pandemic began and the lowest in the record, so that it doesn’t count as “before the pandemic”; including it moves the answer by about two points.

Winters scatter around any benchmark. Taking how far those sixteen winters strayed from it on a log scale, a 90% reference band for last winter’s excess runs from 27% to 73%. It assumes the scatter is roughly normal, so we checked it two other ways: a regression prediction interval gives 29% to 80%, and a bootstrap that resamples the sixteen winters gives 16% to 75%, its low end set by one unusual winter, 2005-06. Every version stays well above zero, and the middle of each is about half. We did not detect a drift within the sixteen winters that would move the benchmark: about -0.6% a year, with a range of -1.5% to +0.3%. Five other checks fall inside the band:

• a model fitted on the same winters, in which complaints rise less than in step with the cold: 52% above;
• a model by month, which lets January and October behave differently: 48%;
• the nine winters since 2003 that were within 5% as cold, which brought between 170,000 and 225,000 complaints: last winter was 55% above their average;
• complaints without those HPD flagged as duplicates (49% of last winter’s), about 157,000: 51% above, and a record every winter since 2021-22;
• complaints per renter household outside public housing, whose tenants complain to NYCHA rather than HPD, with the survey’s count of those households: 31% above.

The last check is the lowest because the survey counts more renter households outside public housing in 2023 than in the winters before the pandemic, and we hold the 2023 count for the winters since. Across the checks the excess runs from 31% to 55%, and the low ends of their uncertainty ranges run from 16% to 29%: all above zero, none a guarantee of a particular floor. It is much the same with the weather at LaGuardia or JFK instead of Central Park, 46% and 45% above, and at all three stations only the last two winters ran above the pre-pandemic range. One more thing changed inside the base winters: from October 2017 the overnight minimum rose to 62 degrees whatever the weather, from 55 degrees when it was below 40 outside (Local Law 86 of 2017). The two base winters after it ran at 0.98 and 0.94 of the benchmark, so the stricter night standard did not by itself lift complaints.

The most detailed check models each day rather than each season. A count model of every heat-season day’s complaints, fitted on the sixteen winters, lets complaints climb more steeply on colder days, carries each cold snap into the two days after it, and allows for the month and the day of the week. It puts last winter’s cold at about 212,000 complaints, so last winter ran 44% above; predicting each of the sixteen winters from a model fitted without it gives a 90% reference band of 17% to 77%. A quasi-Poisson fit has the same coefficients; we have not fitted a negative-binomial model. By that model too, the gap opened after the pandemic: complaints ran 1.29 times the prediction in 2022-23 and 1.44 times in each of the last two winters.

Set against the benchmark, the seasons since the pandemic began run 0.80, 0.85, 1.09, 1.24, 1.25, 1.36 and 1.48 times what their cold predicts. Before, the ratio ran from 0.85 to 1.28, and only 2005-06 passed 1.2; that winter, at 236,000 complaints, was the record until 2024-25.

Cold snaps matter as well as the total. Last winter had eight nights at or below 10 degrees at Central Park, as many as 2003-04, 2004-05 and 2013-14 and one fewer than 2014-15, and its coldest night was 3 degrees; those winters brought 192,000, 206,000, 195,000 and 209,000 complaints.

Possibly, in part, and the counts cannot settle it. The best evidence is the housing survey, for two winters only. Among renter households whose heat broke down, the share that said anyone had contacted 311 was about the same in its two latest rounds, 40% for the winter of 2019-20 and 38% for 2021-22, give or take 4 and 3 points, while the share whose heat broke down rose. That fits more households losing heat; but the survey asks only whether anyone called, not how many times, and complaints count every call.

The channel data can’t tell us whether filing became easier. Most heat complaints now reach 311 online or through its app (Figure 4). The phone’s share fell from more than half before 2021 to about a seventh in 2022 and recovered to about a third in 2024 and 2025, so it does not move in step with complaints. Complaints per building rose too, from 5.9 when the record began to 8.0 last winter, which fits easier complaining as well as colder apartments. The inspections speak to it more directly: the share of complaints inspected while heat was required that turned up a violation rose rather than fell, which is hard to square with easier complaining alone.

Partly. We matched last winter’s buildings to HPD’s register of the buildings it oversees, which gives each one’s number of apartments (98% of the buildings, holding 99.7% of the complaints). The 3,306 buildings with twenty or more no-heat complaints had a median of 50 apartments, against 11 for the rest. Per apartment, they sent 0.77 complaints last winter, against 0.12 for the rest.

The concentration in the second chart is not a few outliers either: leaving out the hundred busiest buildings, the busiest tenth’s share still rose, from 50% when the record began to 57% last winter.

And the repeat buildings are not a fixed list. Of last winter’s buildings with twenty or more complaints, 68% had reached twenty in at least one of the ten winters before, which is no higher than usual: from 2013-14 to 2024-25 the share ran between 69% and 77%, and last winter 1,048 buildings reached twenty with no such winter behind them. What grew is the repeat buildings’ share of all no-heat complaints, from 35% in 2013-14 to 46% last winter.

The busiest tenth of buildings sent 61% of last winter’s no-heat complaints, up from 53% when the record began. That tenth is re-ranked every winter, so it is not the same buildings from year to year, and the pool of buildings with any complaint grew from 32,295 to 38,169; the repeat-building chart in the story is the measure of persistence.

Twenty complaints is an easier mark for a big building than a small one, so we also defined high-complaint buildings by complaints per apartment: half a complaint or more per apartment, in buildings of twenty apartments or more. The pattern is the same. The share of those buildings that were repeats stayed between 69% and 80% each winter, and repeat buildings’ share of all no-heat complaints rose from 26% in 2013-14 to 32% last winter. The apartment counts are today’s, from HPD’s register, applied to every season, so this is a check at current sizes; the register gives a count for 99% of the buildings with ten or more complaints in every season.

January 2026, with 74,168 no-heat complaints to HPD, ahead of January 2004’s 69,553, the previous high for any month. Four of last winter’s eight months, November, January, February and April, were the busiest of their month in the record, and December and March the second-busiest.

Yes, more than twice as many. Last winter HPD closed 13,793 first complaints of no heat with a new heat violation, and 98 more where one was already on file, against an average of about 5,200 new ones a winter from 2003-04 to 2018-19 (complaints flagged as duplicates take the first complaint’s result, so we leave them out; the unit is HPD’s problem record, within half a per cent of distinct complaints last winter). Counting only inspections made while heat was required, the share that turned up a violation rose from about a quarter before the pandemic to about two-fifths last winter, after nearly half in 2023-24 and 2024-25.

How often inspectors got in has swung over the years: between about a tenth and a third of first complaints ended with no access, the most in 2013-14 (36%), and between about two-fifths and two-thirds ended with heat restored by the time HPD reached someone in the building. The night standard changed too: from October 2017 a violation can be written below 62 degrees overnight whatever the weather, where before it took 55 degrees with below 40 outside, and the rise in violations came from 2022-23. Measured the way the Comptroller measured it, against all no-heat complaints, about 3% ended in a new violation before the pandemic and 4.5% last winter; the Comptroller put it at about 3% for 2017 to 2024 and 4.5% in the 2023 heat season. From this winter, HPD must attempt an inspection at every apartment that complains by name, so these counts will change for reasons of their own.

Per renter household, the Bronx stands out. We divide each borough’s no-heat complaints by its renter households outside public housing, since public-housing tenants rarely complain to HPD. Across the city the rate rose from 103 complaints per 1,000 such households in 2003-04 to 141 last winter. The Bronx’s rose from 157 to 267, nearly twice the city’s, and Manhattan’s from 84 to 127. Brooklyn’s ended where it began, at 124, after falling through the 2010s, and Staten Island’s stayed the lowest. The households come from the housing survey, interpolated between its rounds and held at the 2023 count since, so the latest rates lean on a survey three years old.

Cold, but not the coldest. By heating degree days, the season from October 2025 to May 2026 needed 4,791, the most since 2014-15 and more than 23 of the 29 winters before it since 1996-97. The coldest of the thirty, 2002-03, needed 5,292 and the mildest, 2011-12, 3,733 (Figure 9).

Because its tenants report elsewhere. NYCHA asks residents to report heat and hot-water problems to NYCHA itself, through its MyNYCHA app or its Customer Contact Center, so their reports rarely become HPD complaints. Matching last winter’s no-heat complaints by tax lot to the lots in NYCHA’s published residential-address file finds 12 complaints on 8 of its 526 lots; HPD’s own register, which marks NYCHA’s buildings, finds 9. NYCHA’s tenants do call 311: in the survey, about two in five public-housing households whose heat broke down said they had, about as many as in rent-stabilized housing.

Some developments have moved to private management under PACT, NYCHA’s program that uses the federal Rental Assistance Demonstration (RAD). The survey’s codebooks count RAD units as public housing, and NYCHA’s address list leaves most converted developments out, so some of the survey’s public-housing households live in buildings whose complaints may reach HPD; these files can’t tell how many.

NYCHA reports its own heat season instead: 2,240 heat or hot-water outages in 2025-26, 98% restored within 24 hours and in 7.8 hours on average, and, by its October count, 6% fewer outages than in 2021-22 (its June release had put the fall at 7% over five years). An outage at one boiler can leave a whole building cold, so NYCHA’s count and the survey’s share of households answer different questions. NYCHA also says its boilers average about 29 years old, near the upper end of the 25 to 30 years it now expects them to last.

Each of the four records in the story counts something different, so here they are together with their latest readings.

Every survey since 1991 has asked renter households whether their heating equipment broke down for six hours or more. Through 2017 it asked about “this winter”, the one under way or just ended when the interviewer came, so each survey’s winter depends a little on when its interviews ran. From 2021 it asks about one fixed heat season: October 2019 to May 2020 for the 2021 survey, and October 2021 to May 2022 for 2023. The 2021 survey was also redesigned, with a new sample, new weighting, new question wording and interviews on laptops rather than paper, and missing answers were filled in by the Census Bureau’s standard methods rather than left out; HPD’s 2021 report says the heating measure changed. So the two stretches are separate series: the third chart breaks between them, and we compare years only within one.

One change sits inside the old stretch. In 2017 the survey began to use the state’s rent registrations to code about 62,000 apartments registered as permanently exempt as unregulated rather than rent-stabilized, and HPD cautions against comparing the stabilized population across that change. Giving those apartments any breakdown rate from none to 30% would move the 2017 share by at most about a point for rent-stabilized rentals and two for market-rate rentals, which changes none of the comparisons here; the shares for all renters and for public housing don’t depend on it.

We leave out households that did not answer (through 2017) or, from 2021, did not live there that winter, as the Census Bureau’s published tables did, and our counts match those tables to the household. The unanswered share was sizeable in some years: 31% of renter households in 2008, against 6% in 2017. Margins cover sampling, not the answers that are missing, and those who didn’t answer may differ from those who did; the setback from 2008 to 2011 compares a year when nearly a third didn’t answer with one when about a seventh didn’t, so read it with care. From 2021 the share is of households that lived there that winter, which is not the same as every household in the city that winter: those who moved out before the survey are not in it. HPD’s published shares for 2021 and 2023 divide by every household, including those that moved in after that winter, so they run lower than ours: 17% and 5% for rent-stabilized and private unregulated rentals in 2021 (ours 19% and 7%), and 20% and 9% in 2023 (ours 22% and 11%). The counts of households with a breakdown agree.

From 2011 the files carry 80 replicate weights, which give the margins directly. Before that we use the Bureau’s generalized variance formula with the parameter its accuracy statements give items like this one, which tend to apply to every apartment in a building: for 1991 to 1999 the statements name heating breakdowns among them, and for 2002 to 2008 we take the largest of their housing parameters, to be safe.

A change is called a rise or a fall only where its 90% range clears zero. From 1991 to 2017 the falls for all renters (24% to 13%), rent-stabilized (29% to 14%) and private unregulated rentals (14% to 8%) clear it, and so do the setbacks for all renters from 2002 to 2005 and from 2008 to 2011. Public housing’s rise from 19% in 2002 to 29% in 2017 clears it (5 to 16 points), and so does the doubling of its gap with private unregulated rentals, from 10 points in 1991 to 21 in 2017; its change from 1991 to 2017 alone does not. Between 2021 and 2023, the rises for all renters and for private unregulated rentals clear zero, while those for rent-stabilized rentals and public housing (35% to 39%) are within their margins, and so is the change in the gap, which stood at 27 points in 2023.

Gaps between groups within one survey come from the replicate weights from 2011, so the two groups’ covariance is kept; the 1991 gap treats them as independent, an approximation. Between surveys we treat the estimates as independent, which overstates the margin a little where samples overlap, as the 2021 and 2023 ones do. In 2023 public housing’s share was 3.4 times the market-rate share, with a range of 2.9 to 3.9 times. The files differ in one more way through 2008: the 2002 to 2008 files (the Census Bureau’s 2008 longitudinal release) use a revised recode that ranks rent stabilization ahead of HUD regulation, and the earlier files the original one; public housing and private unregulated rentals are coded alike in both, but the stabilized line may shift a little between 1999 and 2002 for that reason. Groups follow the survey’s control-status code: public housing, rent stabilized (built before or after 1947), and private unregulated rentals (“other rental” before 2021). All renters also include rent-controlled, Mitchell-Lama and other regulated rentals.

A measure of how much heating a day needs. Take the day’s mean temperature at Central Park; every degree it falls short of 65 Fahrenheit adds one degree day, so a day averaging 30 adds 35, and a day at 65 or warmer adds none. Adding up the days from October to May gives a season’s total, which compares winters with one another rather than measuring what any apartment felt.

Complaints. HPD’s Housing Maintenance Code Complaints and Problems file holds every complaint the agency has logged since February 2003, each with one or more problems. A no-heat problem is one whose code begins NO HEAT (no heat, or no heat and no hot water), in either of the file’s two category schemes (HEATING until March 2014, HEAT/HOT WATER after); hot water alone, heat left on in summer and the old scheme’s boiler and radiator codes (about one complaint in a hundred) are left out. Each no-heat problem counts once, by the month its complaint was received. A heat season runs from October to May, as the law does. The duplicate flag and its kind (building-wide or one apartment) come from the same file. Every count was made on the City’s server (the queries are saved with the date retrieved), so the files hold no addresses.

The cold. Heating degree days are NOAA’s monthly figures for Central Park, published in Celsius degree days on a base of 18.3 C and multiplied by 1.8 for the Fahrenheit, base-65 figure; NOAA’s daily file gives the same seasons within two per cent. The benchmark is each season’s degree days times the complaints per degree day of the sixteen seasons 2003-04 to 2018-19 taken together. Its 90% band is empirical, from the scatter of those seasons around it; beside it we give a prediction interval from the log-log regression and a bootstrap that resamples the sixteen seasons (20,000 draws, a fixed seed). LaGuardia’s and JFK’s monthly figures give the same benchmark from two other stations. The daily model is a Poisson regression (a quasi-Poisson model, which allows for over-dispersion, has the same coefficients) of each heat-season day’s no-heat complaints on linear splines of the day’s mean temperature and the day before’s (bends at 25, 35, 45 and 55 degrees) and of the mean two days before (bends at 35 and 50) and of the night’s low and the night before’s (bends at 10 and 20), with a term for each month and each day of the week, fitted on every day of the sixteen seasons; its band comes from predicting each of those seasons from a model fitted without it. The two models are fitted on the same seasons: log complaints on log degree days by season, and log complaints on log degree days with a term for each month, summed over the season’s months. The matched winters are those sixteen seasons within 5% of the season’s degree days. The per-household check divides by renter households from the survey, interpolated between survey years and held at the 2023 count after.

Buildings. Buildings are HPD’s building records. For each season the server counts every building’s no-heat complaints and returns how many buildings had one, two, three and so on, which gives the busiest tenth’s share; a second query lists the buildings with ten or more, used only to count how many buildings with twenty or more had reached twenty in any of the ten seasons before (from 2013-14, the first season with ten before it). Apartments per building are the legal Class A units in HPD’s register. How a complaint was closed comes from HPD’s closing note: a violation issued (or already on file), inspected with no violation, heat restored by the time HPD reached someone, no access, or still open; duplicates take the first complaint’s result, so the rates count first complaints only. A second definition of a high-complaint building uses complaints per apartment (half or more, in buildings of twenty apartments or more), which needs ten complaints and so is complete in the same query. Boroughs are rated per 1,000 renter households outside public housing, from the survey. No building is named, mapped or published. Public-housing complaints are counted by matching last season’s complaints by tax lot against NYCHA’s residential addresses, and by the register’s NYCHA flag.

The survey. The occupied-unit files of the Housing and Vacancy Survey for 1991, 1993, 1996, 1999, 2002, 2005, 2008 (the Census Bureau’s 2008 longitudinal release for the last three), 2011, 2014, 2017, 2021 and 2023, read with the record layouts saved beside them. Renter households are those that rent or occupy without paying rent; a breakdown is one of six hours or more. Margins are 90%, as the survey question above describes.

Code and checks. One script, analysis.py (linked below), produces every figure and chart from these files. The dataset cards behind it (hpd_heat_complaints, nyc311_heat, noaa_central_park, nychvs_history, nychvs_2023) run the checks quoted in the second question: no month missing, HPD against 311 each season, the City’s published totals, the Census Bureau’s and HPD’s published survey tables, the Bureau’s variance parameters in its saved accuracy statements, and the two weather products against each other.