When nature calls, where should New York answer?

Archived version of 24 Sep 2026 (the merged walkthrough), kept for comparison. Some figures and claims are superseded — current figures are in the Walkthrough.

Merged walkthrough (24 Sep)

0 — The situation

• Local Law 58 of May 2025 sets a target of 2,120 public restrooms by 2035. The City's register lists 975 operational today, so the gap is 1,145.
• The City's siting strategy under the law is due on 1 Nov 2026.
• Earlier laws asked for sites without funding them: Local Law 114 of 2022 wanted a feasible site in every ZIP code area, and Local Law 92 of 2025 reports on the projects at those sites. Team research · Steph, Candy

New York has about a third of the public toilets per resident of the best-served cities. Counted the same way, as government-run toilets on streets, in parks and plazas and at transit stops, it has 9.0 per 100,000 residents. Team research · Shawn (idea)

Hong Kong's list excludes park toilets run by another department; with an approximate count of them it is about 18.4.

Toronto and Berlin are close to New York, so the gap is with the leaders, not with every large city.

The City has already started a 17-unit modular pilot:

Sources: PIX11, Aug 2026, for the deadline; NYC Mayor's Office release of 16 Sep 2026 for the pilot and its sites Team research · Steph, Giorgia.

The City has four decisions to make: where to put restrooms, when they need to be open, what to do first, and how to measure the pilot it has already started. The sections below take each in turn.

1 — Where

Data cleaning, negative binomial regression, out-of-sample validation, hot-spot mapping

The signal is 311 "Urinating in Public" complaints. Nobody records "needed a toilet and could not find one". This is the closest thing that is dated and mapped: 7,926 complaints from 2020 to 2026.

Clean it first. The top 1% of addresses produce 22.3% of complaints. In East Harlem, 223 of 384 complaints came from five addresses on one block.

Keeping complaints in public places leaves 4,706, and counting each address at most once a month leaves 3,629. Otherwise five addresses on one block would set East Harlem's rank.

Adjust for crowds. Busy places generate more complaints. For each of 197 neighbourhoods, the model predicts the complaints it would normally produce, given its subway riders, jobs, hotels, residents, income, age mix and borough.

Observed ÷ expected is the unmet-need ratio. Brighton Beach, for example, had 24 complaints against 6.2 expected: a ratio of 3.85×.

Test it against a one-line rule. The rule ranks neighbourhoods by subway ridership alone. The model is fitted on 70% of neighbourhoods and scored on the unseen 30%, over 200 random splits.

On rank agreement with actual complaints, the model scores 0.788 against the rule's 0.703. It wins in 96% of 200 splits.

Shaded by observed ÷ expected complaints. Grey means the two match; stronger colour means more complaints than the area's crowds explain.

The answer: 29 neighbourhoods sit at 1.5× or more of their expected complaints. Six clear 0.95 probability that their true complaint rate is above 1.5× expected: East Harlem (North), Astoria (East)–Woodside (North), East Elmhurst, Williamsbridge–Olinville, Brighton Beach and Midtown–Times Square. Interactive map.

That confidence is about excess complaints. Re-ranked on phone-filed complaints alone, 4 of the six fall out of the top 25, so it is weaker evidence of excess need.

Where exactly: the complaint map shows the blocks. Within each of the six, complaints cluster in a few hot spots. The table shows the largest one, whether it sits beside a subway station, and which agency controls the ground.

In Midtown the complaints cluster at station exits. In East Elmhurst and Williamsbridge it sits at playgrounds whose restrooms are locked at night. Astoria's hot spot rests on 3 addresses and stops in 2022, so it may reflect one persistent caller or a problem that has ended.

A hot spot counts as beside a station when its centre is within 250 m of one. The owner follows a stated rule: beside a station, MTA or DOT; within 60 m of a park, Parks; otherwise DOT. It starts a discussion; it does not settle one.

The pilot mostly misses the need. 1 of 17 units sits in one of the six, at Northern Blvd and 54 St, on the area's boundary. 3 of 17 are in the 29, and the typical site's neighbourhood ranks 82nd of 197.

The ranking's value is in where the next units go. The pilot is already sited; the LL58 strategy decides the next 1,145.

Caveat: the 1.5× cut-off is a budget line, not a natural boundary.

2 — When

Hours parsing, walking distance by hour, complaint timing

Access collapses in the evening. Counting only restrooms with genuine posted hours, 18% of residential land is within a five-minute walk of an open restroom at 3pm. At 9pm it is 2%.

Share of each neighbourhood within a five-minute walk of a restroom open at that hour, counting only restrooms with genuine posted hours. Interactive map.

On a weekday evening, only 57 restrooms are open at 9pm. The chart shows how far a rider leaving one of the 50 busiest subway stations must walk to reach a restroom that is open at that hour. Team research · Steph

The true summer closing time is unknown. 641 of the 975 operational restrooms show the same placeholder hours, "8am-4pm, Open later seasonally". The main 9pm row assumes park restrooms close by 9pm.

"Open 24 hours" means a gas station. The register lists 8 facilities as open 24 hours. All 8 are parkway gas stations or a parkway snack bar. Team research · Steph

Among the six priority areas, the evening gap is worst in Astoria and smallest in Midtown. The table gives the median walk at 9pm from where complaints are filed to the nearest open restroom.

Astoria's figure rests largely on the thin hot spot flagged in section 1, which holds 36 of its 59 complaints and none after 2022. Williamsbridge–Olinville, at 36.6 min, is the worst case whose complaints run through 2026.

Midtown is the best served of the six at 9pm, so a shortage of open restrooms does not explain its excess complaints. Nightlife or street homelessness may; the Limits tab keeps this open.

Complaints lean toward the evening and the night. 30% are filed between 6pm and midnight. Against 311 dirty-condition complaints, which reflect the same calling habits, urination complaints run 1.4–1.9× as high between 6pm and midnight and 1.4–2.3× between midnight and 5am.

That pattern fits closing times. It fits nightlife and alcohol just as well.

Why complaints concentrate here: access is the leading hypothesis. Outside evidence is mixed, and our data cannot confirm it, so it should be tested.

The outside evidence comes from San Francisco, not New York. There, 13 new staffed Pit Stop restrooms were followed by 12.47 fewer feces reports a week, p = 0.0002, but three Pit Stops given longer, 24-hour service saw 12.00 more a week, p = 0.0016, which the authors call unexpected (Amato et al., 2022). A 2025 follow-up covering 2009–2022 found that "the effect appears to be small" (PLOS ONE, 2025). Team research · Steph

In New York, the largest hot spots in Williamsbridge and East Elmhurst sit 47 m and 126 m from park restrooms that are locked at night.

Our data cannot confirm it citywide. Once crowds are counted, no measure of supply predicts complaints: not the number of restrooms, not the number open at 9pm, not the share in parks. Most park hours are placeholders, counts are thin, and restrooms tend to be built where problems already are, which can hide an effect.

The hours gap in the data is fixable. Paris publishes one open dataset of every public toilet with its status and opening hours.

Whether night closures drive complaints is untested, and cheap to test. Section 4 sets out a test at the two named park restrooms.

Caveat: complaint times are filing times, not the time of the incident.

3 — What first

Panel data, facility fixed effects, placebo test, delivery times, cost analysis

The existing stock is deteriorating. This is the strongest result here. NYC Parks inspectors grade every comfort station on a schedule: 31,768 inspections.

Comparing the same January–June window each year, the unacceptable share was flat for three years, then jumped in 2025–26.

That is 9.4% → 20.1% since 2022, and 8.5% → 20.1% since 2024. The 2024 and 2026 figures rest on 743 and 693 rated inspections.

• It is the same buildings getting worse. With facility fixed effects, 2026 is +11.4 pp above 2024.
• It does not look like stricter marking. The 7 inspectors who rated restrooms in both years marked litter no worse, 0.016 → 0.011, while structural failures rose from 0.044 to 0.210.
• A field audit finds the same problem on the ground. The City Council's random "Good to Go?" audit, Aug–Sep 2025, found 36 of 337 rooms locked during posted hours, all in parks, and 129 of 301 open rooms missing a basic necessity. It is a single snapshot, so it confirms the level, not the trend. Team research · Steph, Candy
• 80 of 736 Parks comfort stations are under long-term closure, 46 of them for repairs.

Each option is priced per area, one facility each, like for like:

The lowest-cost option in each of the 29 places adds up to $24.1M in capital plus $0.51M a year to run the longer hours:

An area is classed "close early" when the median posted day of its restrooms is under 9 hours, or none is open to 22:00. All 10 have at least one restroom with real posted hours, and 9 of 10 keep the class when placeholder-hours restrooms are left out.

Dividing each option's yearly cost by its area's excess complaints, the lowest ratio is longer hours in East Harlem (North), at $5,199 per excess complaint a year. Next is a modular unit in Astoria (East)–Woodside (North), at $14,727. This ranks cost against the size of the problem; it is not a measured effect.

Longer hours are the cheapest option to test, not a proven fix. They beat a modular unit only if they are at least 63% as effective as one. The median costed area, of 27, has 1.8 excess complaints a year, too few to measure an effect area by area.

Repair is a service-quality priority, not a proven way to cut complaints. The decay is citywide rather than concentrated where complaints are high: in the 29 priority areas, 15 of 72 inspected comfort stations were rated unacceptable in 2026, against 116 of 507 elsewhere. Long-term closures run slightly higher there, 14 of 100 against 66 of 636.

Repair now, and test longer hours alongside. The stock is measurably decaying, and a reconstruction costs well under half as much as a new permanent building. It also delivers sooner: a median 4.5 years from design to completion, against 7.4 years for a new building, on 70 and 38 completed projects, p < 0.001.

A closed restroom serves nobody, while a reopened one serves whoever is there.

Caveat: the hours cost is an assumption. Hours stay the lowest-cost option only while staffing costs less than about $100/h, under this one-facility-per-area convention.

4 — How to measure the pilot

Power analysis, service metrics, randomised staggered trial

An impact evaluation of the 17 pilot units would be underpowered. Only 20 complaints a year are filed within 400 m of all 17 sites combined, during the pilot's opening hours. A before-and-after comparison with matched areas could detect only a 73% drop in complaints, or 44% using urination summonses.

With 1 of 17 units in a high-confidence area, the pilot cannot test the ranking either.

Judge the pilot on service: the hours it is actually open, not the units installed. Team research · Steph, Candy

• Cost per verified open hour divides total cost by the hours each unit was found open and working. Against scheduled hours the pilot costs $42.98 per unit-hour: $4M ÷ 17 units × 15 hours × 365 days. The Council audit found 36 of 337 rooms locked during posted hours, so scheduled hours overstate service.
• Stage gates. Verify that a site is really underserved, confirm the unit is open and working, then scale up. Only after that should a site get a permanent building.
• Data to require from the operator, site by site and week by week: open and closed timestamps, uses, faults, time to repair, and cost. Washington DC's pilot law, D.C. Act 25-172, already requires monthly reports on use and misuse.

Test impact where it can be seen: longer evening hours. Start with the two park restrooms beside the largest hot spots, Gun Hill Playground in Williamsbridge and Gorman Playground in East Elmhurst, and compare them with matched restrooms. Then extend the same design to the 10 areas that close early.

The trial opens every restroom in the 10 areas, not the one per area costed above, so it tests the strongest version of the option.

Caveat: summonses partly track policing, so they cannot rank places. Within one area over time they can measure change, unless longer hours also change patrols, which the alcohol-summons placebo, a count of alcohol summonses that longer hours should not change, is there to catch.

5 — What we recommend

6 — How we know it holds up

• The model scores 0.784–0.791 across six random seeds. A decision tree scores 0.720.
• Adding 1,886 chain cafés and fast-food outlets as informal supply leaves the ranking unchanged, at 0.997.
• Each of the six high-confidence areas was re-scored with itself left out, and all held.
• The weak points: the ranking does not fully survive a phone-only check; the ratio may partly track street homelessness or nightlife; and no New York estimate yet shows that a restroom reduces the problem. The limits section below sets out each one.

Merged version: limits and open questions

This tab collects every weakness in one place, starting with the questions we have not yet answered.

Open questions: not yet resolved

These are the four hardest objections to the analysis. None has a full answer yet.

1. The need ratio may partly map street homelessness or nightlife. The model adjusts for subway riders, jobs, hotels, residents, income, age mix and borough. It has no control for shelters, unsheltered counts, bars or late-night licences. Midtown–Times Square is the sharpest case: it is the best served of the six at 9pm, so its excess complaints need an explanation other than supply. A neighbourhood with many people living on the street, or a busy bar strip, could show a high ratio for reasons a new restroom would only partly fix. Next step: add DHS shelter counts by neighbourhood as a control and re-rank.
2. The confidence figures and the phone-only check point in different directions. East Harlem (North) and Astoria (East)–Woodside (North) both have a posterior probability above 0.999. Yet if we re-rank using only complaints filed by phone, East Harlem falls to 41st, Astoria to 121st, Williamsbridge–Olinville to 50th, and Midtown–Times Square from 17th to 42nd: four of the six. The two results fit together once you see what the posterior assumes: it measures how sure we are given that complaints measure need. The phone-only check questions that assumption. The right reading is "very sure there are excess complaints here", which is weaker than "very sure there is excess need". Of the high-confidence areas checked, Brighton Beach and East Elmhurst hold under both.
3. The trial outcome is the one we rejected for ranking. Across neighbourhoods, urination summonses correlate 0.95 with alcohol summonses, which means they map policing more than need. We therefore did not use them to rank places. In a before-and-after design each area is compared with itself, so constant enforcement cancels out. But if opening restrooms for longer changes patrol patterns, it does not cancel. The alcohol-summons placebo would catch a general shift in patrolling, though not one aimed specifically at restroom areas.
4. Do night closures cause the complaints? Access is the leading explanation: outside evidence from San Francisco supports it, and the largest hot spots in Williamsbridge and East Elmhurst sit beside park restrooms locked at night. But no measure of supply predicts complaints once crowds are counted (§12). Next step: the evening-hours test at Gun Hill and Gorman Playgrounds, with matched controls.

1. The ranking orders places; it does not sort them into categories

Complaint counts are thin, so there are two things to check.

• Is the variation real? Yes. The spread across neighbourhoods is wider than chance alone would produce.
• Are the worst places a separate group? No. Suppose every neighbourhood genuinely differed by exactly the spread we estimated. Simulations of that city throw up 33.8 neighbourhoods above 1.5× on average; we found 29. Need shades continuously from the top of the list downwards, so the 1.5× threshold is a budget decision, not a discovery.
• How sure are we about one place? The posterior probabilities (Walkthrough, "Where") answer this. Six neighbourhoods clear 0.95; East Flatbush–Rugby just misses at 0.948. Each result held when the neighbourhood was left out of the model and the prior. A formal correction for testing all 197 places at once would clear none of them. So the honest claim is about where to look and where to spend first, not a proven fact about any single place.

2. We never validated the thing we actually deliver

The model predicts complaints well: 0.788 against the one-line rule's 0.703, scored on neighbourhoods it had not seen. But what we deliver is the gap between prediction and reality, and nothing can validate that gap, because "true unmet need" has never been measured. The better the model predicts, the less gap is left to rank. Excess-mortality estimates work the same way, but they rest on counts in the thousands; ours are in the tens.

3. We cannot yet show that a restroom reduces street urination in New York

We tried two causal designs. Both failed:

• Closures. The inspection records flagged as "closures" turned out mostly to log construction activity, not real closures, so there was nothing to measure.
• Winter closures. Seasonal restrooms shut every winter while year-round ones stay open, which looks like a natural experiment. Urination summonses near seasonal sites fell (IRR 0.624). But the placebo, alcohol summonses, fell almost as much (IRR 0.699). Parks empty out in winter, so we had measured the season, not the restroom.

So this project contains no New York effect estimate, and no return-on-investment figure. Published evidence does exist from elsewhere. In San Francisco, Amato et al. (BMC Public Health, 2022) found −12.47 feces reports a week after 13 new Pit Stop restrooms opened. A 2025 follow-up in PLOS ONE found "the effect appears to be small". What does not exist is a US cost-benefit study. Our break-even figures show how large an effect would have to be to justify the spend; they are not a measured effect.

4. The cost comparison depends on assumptions

• Every option is priced for one facility per area, like for like. On that basis the $24.1M capital figure has a running cost of $0.51M a year for the longer hours. The trial is costed separately because it opens all 51 restrooms in its 10 areas: $2.6M a year.
• The staffing cost ($35/h) and the extra hours per day (4) are assumptions, not sourced figures. Extending hours stops being the cheapest option at about $100/h.
• If every restroom in an area needed its own extra staffing, East Harlem (North) would fall from 1st to 9th in cost per excess complaint.
• Installing a unit in all 29 places would cost $34.8M (29 × the $1.2M Parks pilot price). Of course the cheapest-fix mix costs less than that; the comparison is true by construction and is not a finding.
• The modular price ($1.2M, from NYC Parks' Portland Loo pilot) is for a unit connected to water and sewer. The 2026 pilot's $235k per unit-year is a service contract (units plus cleaning), so the two figures are not comparable.

5. Complaints measure complainers

• Across community boards, complaint counts correlate 0.31 with the share of reports filed online.
• On phone-only reports, four of the six high-confidence areas drop sharply, Midtown–Times Square included (17th → 42nd; see the open questions above).
• The person who complains and the person who needed a toilet are different people. To some extent we are measuring annoyance.
• We tested police summonses as an alternative: they are 27,384 records and far denser. They correlate 0.95 with alcohol summonses, so they map enforcement. We kept complaints.

6. The parks explanation is inferred, not measured

824 of 1,066 restrooms are in parks, and capital money follows that existing stock: an area's existing restroom count is the strongest predictor of a restroom capital project, with an odds ratio of 4.72 per log-unit of restroom count. Measured need points the right way but is not significant in any version: 1.63, p = 0.075.

Once crowds are accounted for, supply shows no link with complaints, and the share of restrooms in parks does not predict them either (§12). Park-sited and street-sited restrooms were never compared directly, because there are too few street-sited ones to compare.

7. The trial is blunt

Measured on 311 complaints, the trial could only detect a change of 156%, or 433% using phone-filed reports alone. Measured on summonses, it can detect 37%. That is enough to catch a large effect but not a modest one, and the summons outcome carries the policing caveat in open question 3.

8. The deterioration finding is strong, but not airtight

• The within-inspector test rests on 7 people. Standard errors clustered on seven inspectors are unreliable.
• A gradual shift in inspector emphasis cannot be ruled out. The January–June rate was flat from 2022 to 2024 (9.4%, 9.0%, 8.5%) and then jumped (11.3%, 20.1%), which could fit a change in how inspectors rate as well as real decay. 2024, our comparison year, is the lowest of the five. New inspectors fail restrooms less often, not more. We found no published change to the inspection method, but that does not prove there was none.
• Park-wide condition makes a poor placebo. Parks without a comfort station improved while parks with one got worse, so the citywide average hides two opposite movements. Site condition also moves with restroom condition. We rely instead on the internal placebo: the same inspector on the same visit marked litter as flat while structural failures rose.
• 2026 is a part year. Every year is compared on the same January–June window.
• It measures condition, not siting. The siting question still rests on complaints.
• Component-repair costs are too thin to quote. Only 5 projects in the capital tracker are component work.

9. Data problems worth knowing

• Placeholder hours. 525 of the 843 year-round restrooms have the same text: "8am-4pm, Open later seasonally". The City's own register cannot say what is open at 9pm.
• A stale register. Its rows were last updated on 27 June 2025. The City Council's dashboard says 973 are operational; the register gives 975, and we use the register. The dashboard's "49.42% within a five-minute walk" has no published method, so we do not use it.
• The capital tracker repeats rows. Its entries cover 191 restroom projects, and bundled entries repeat costs. Many projects give only a cost band, and those are the expensive ones, which is why the pooled median is $2,331,000.
• Retired 311 complaint types return zero rows without any warning.

10. Datasets that looked useful and were not

11. A station-level screen did not beat ridership

A screen of 424 subway station complexes scores each station on its ridership, its walk to the nearest listed restroom and whether any restroom is within 500 m. Tested against where complaints actually fall, it ranks stations worse than ridership alone: 0.467 against 0.624 out of sample, worse in 100% of 200 splits.

The restroom-distance part adds nothing once ridership is known. Each doubling of the distance changes complaints by a factor of 0.94 [0.86–1.03], which is indistinguishable from none.

The station list remains a useful inventory of busy places, and the hub figures in the Walkthrough describe access. Locating hot spots is done with the complaint map.

12. The supply-mechanism tests came back null

Evening supply was added to the ranking model to test whether restrooms in the wrong places, or closed at the wrong hours, explain complaints. None of the measures carries a signal once crowds are counted:

The best of these improves out-of-sample accuracy by +0.004. Areas where coverage collapses most after dark do not have a more evening-weighted complaint mix either: odds ratio 1.00 [0.92–1.09].

Three reasons stop this from showing that restrooms do not matter:

• The hours data is mostly placeholder. 641 of the 975 operational restrooms show the same generic hours, so evening supply is measured with error.
• The counts are thin. Only 57 restrooms are open at 9pm on a weekday, so there is little variation to test.
• Placement may run in reverse. Restrooms tend to be built where people and problems already are, which pushes any cross-sectional estimate toward "more supply, more complaints".

13. The pilot's impact cannot be read from complaints

Only 20 complaints a year are filed within 400 m of all 17 pilot sites combined, during the pilot's opening hours. A matched before-and-after comparison could detect only large drops:

All three are weaker than the hours trial's 37%. That is why the Walkthrough judges the pilot on service and puts the impact test in the hours trial.

14. The peer-city comparison is approximate

• Each city's list covers different things. New York's register is mostly park comfort stations, Paris's list is mostly street kiosks, Hong Kong's leaves out park toilets and San Francisco's leaves out some street toilets. Every record was sorted into one rule by hand, so treat the table as three groups, not a precise league table.
• New York's 9.0 counts park, plaza and transit restrooms only. Counting every facility on the register, libraries included, gives a higher figure that is not comparable with the other cities.
• The exact rank is not robust. New York is 8 of 8 on the core rule, but Toronto, Berlin and Hong Kong's list are within 2–12% of it. The gap with Paris, Zurich and San Francisco holds under every variant tested.
• Populations come from each city's own statistics office, so years and boundaries differ slightly.

Every dataset and script is listed in the Data & code tab.

Archived version of 22 Sep 2026, kept for comparison. Some figures are superseded — current figures are in the Walkthrough.

Walkthrough (22 Sep)

The whole project, step by step, with the arithmetic shown at each stage.

The answer, up front: twenty-nine neighbourhoods need attention. Nine need a building, ten need only to stay open four hours later, and the rest need repairs to what is already there. Before committing the capital, spend $1.9M finding out whether any of it works.

The situation

• New York has 975 operational public toilets, of 1,066 on the city's register. (The City Council's dashboard says 973; we use the register we can audit.)
• A 2025 law — Local Law 58 — sets a target of 2,120 by 2035 and requires the city to publish a plan for getting there.
• 2,120 − 975 = 1,145 more facilities implied.
• The city has no published method for deciding where they go.
• Last time it tried: 20 automatic toilets bought, 5 installed, 15 left in a Queens warehouse for over a decade. The units were paid for; the city could not settle where to put them.

• For scale: a median toilet project costs $2,331,000 to build.

Our question: where should the next ones go?

Four bodies would act on the answer, and each controls a different decision:

The measurement problem

• Nobody records "person needed a toilet, could not find one."
• New Yorkers report street problems by calling 311, the city's non-emergency service line. Every call is logged with a date, a location and a category, and the records are public.
• The closest observable thing to our problem is therefore the category "Urinating in Public" — roughly eight thousand calls over six years.
• So that becomes the signal — imperfect, but real, dated and mapped.

Step 1 — Clean the signal

Descriptive statistics, data quality

Raw complaint records, filtered down:

Why that last step matters:

• The top 1% of addresses produce 22.3% of all complaints.
• East Harlem ranked worst in the city — but 223 of its 384 complaints came from five adjacent addresses on one street.
• Those five addresses are one block, reported repeatedly by the same handful of people. The count describes a dispute on one street, not a shortage across a neighbourhood.
• After collapsing repeats, the worst areas become Midtown, the Upper West Side, Chelsea, the Village, Elmhurst, Jackson Heights. Busy, plausible.

Finding 1: the raw data points at the wrong places. A handful of addresses dominate it, so any ranking built on uncleaned counts sends the city's money to whichever streets have the most persistent callers. Skipping this step changes the answer completely.

Step 2 — Account for crowds

Multiple regression (negative binomial), dummy variables

• Complaint counts rise with the number of people present, so a raw count mostly measures footfall.
• To separate the two, we estimate how many complaints a place would be expected to produce given how many people are there.

What goes into the estimate:

Everything is public data. The full list, the pull dates and the code are linked at the foot of this page.

• Method: a negative binomial regression — a standard model for counting things that cluster unevenly.

What the model learned. Fitting it to the whole city gives the size and direction of each effect:

The first three are unsurprising: more people present, more complaints. The last one is not, and it becomes step 6.

• Output: for every neighbourhood, an expected number of complaints — the number a place that busy would normally produce.

Step 3 — Find the places that break the pattern

Residual analysis

The calculation, for each of 197 neighbourhoods:

observed complaints ÷ expected complaints = the unmet-need ratio

Above 1.0 means more complaints than predicted, below means fewer. This is the measure used for the rest of the analysis.

Top of the list — read the first row as: Brighton Beach recorded 24 complaints where the model expected 6.3, which is 3.8 times as many.

Finding 2: the answer is not the obvious one. Ranking by raw complaints points at Midtown — already the best-served part of the city. Ranking by unexplained complaints points at ordinary outer-borough neighbourhoods.

Step 4 — Check the model is worth having

Train/test validation, benchmark comparison, decision tree

Before trusting the model, we check it against a deliberately crude alternative.

• The test: each rule predicts how many complaints a neighbourhood should have. Rank the 197 by that prediction, rank them by the complaints they actually recorded, and measure how closely the two orders agree. The measure runs from −1 to 1; 0 means no relationship, 1 a perfect match, and around 0.7 is strong agreement.
• The crude rule: rank neighbourhoods by subway ridership alone, one line of code.
• A more targeted version: count ridership only at stations with no toilet within 400m. The thinking is that crowds who already have a toilet nearby are served, so unmet demand should sit at the stations without one.
• Head to head on the same data, plain ridership scores 0.713 and the targeted version 0.487.
• Narrowing it made the prediction worse, and the reason is informative: a huge station with a toilet beside it still has huge crowds and still generates complaints. Filtering those stations out threw away real demand. Proximity to an existing toilet turned out not to matter much — an early hint of something step 6 finds directly.
• The full model, tested on neighbourhoods it had never seen, scores 0.788 against the crude rule's 0.703 on the same test.
• Those two figures come from hiding 30% of neighbourhoods, training on the other 70%, and scoring the predictions against the hidden ones — repeated 200 times with different random selections. The model won 189 of them.

Finding 3: the model beats a one-line rule — 0.788 against 0.703, winning 96% of 200 trials.

Step 5 — The answer

Spatial analysis and mapping

Run the validated model across all 197 neighbourhoods and this is the output: a map of unmet need, coloured most strongly where complaints most exceed what the area's crowds explain.

Each neighbourhood is shaded by how many complaints it had against how many the model expected for a place that busy. Grey means the two match. Shaded areas had more complaints than their crowds account for — which is why a busy district with a high complaint count can still read grey, while a quiet one with far fewer can be the darkest on the map.

• The map is labelled by borough, with Times Square marked for orientation — Manhattan is the narrow island at the centre.
• Grey covers 150 of the 197 neighbourhoods — those within 25% of what their crowds predict. The 47 coloured ones are more than 25% above it. (78 exceed prediction at all; the map shades only the clearer cases.)
• Of those, 29 sit at 1.5× or above — half again as many complaints as predicted. That threshold is the working shortlist for everything that follows; the eight most strongly coloured areas on the map are its top end.
• The 1.5× line is a budget decision, not a natural boundary. Need shades continuously; nothing in the data marks where "normal" ends. Draw the line elsewhere and the shortlist grows or shrinks smoothly.
• The remaining steps take up a different question: why these places, and what would fix them.

Step 6 — Why the gap exists: the toilets are in the wrong places

Logistic regression

Each map highlights the top thirty by each measure — the supply map shows a few more, because many areas tie. On the left, those with the most complaints beyond what their crowds explain; on the right, those with the most public restrooms. Only five neighbourhoods appear on both.

• 824 of the city's 1,066 catalogued public toilets sit inside parks. (973 of the 1,066 are currently operational; the rest are closed or seasonal.)
• Complaints cluster on streets — transit, commercial strips, nightlife.
• We modelled which neighbourhoods receive toilet construction money. Parkland predicts it clearly. Measured need does not come through at all clearly — its effect could be substantial or could be nothing, the data cannot separate the two.
• And once you account for how busy a place is, how many toilets an area has tells you nothing about its complaint count (rate ratio 1.001, where 1.000 means no relationship at all).
• Toilets get built where problems already are, so a snapshot cannot separate cause from effect. What it does show is that the current stock is not visibly reducing the problem anywhere — which sent us looking at where the toilets are, and when they are open.
• Midtown has the highest complaint count in Manhattan and has never received a Parks toilet project — because Midtown has almost no parkland.

Finding 4: the funding route favours parkland. Restroom money moves through Parks capital projects, and it lands where parkland is. Midtown–Times Square is among the highest-complaint areas in the city and has never received one; it has almost no parkland. Three pieces of evidence point this way, but we never compared park-sited against street-sited facilities directly — see the second tab.

Step 7 — Why the gap exists: they close too early

Spatial coverage analysis, time-of-day comparison

Share of each neighbourhood lying within a five-minute walk of a restroom that is open at that hour, counting only the 316 facilities whose opening times are genuinely recorded. Across the city the covered area falls from 24% at 3pm to 3.5% at 9pm, and 154 of 197 neighbourhoods retain almost none.

• Peak complaint hour: 6pm. That is the hour availability collapses.
• Only 8 facilities citywide are open 24 hours.

Is the 6pm peak real, or just when people are free to phone? 311 records when someone called, not when the incident happened, so we checked it against a different complaint type filed by the same public through the same channel.

If the evening peak were simply calling habits, both would peak together. They do not, and urination complaints are 1.7 times more evening-weighted. Some shared daily rhythm exists, but the peak itself is real.

Finding 5: supply disappears exactly when demand peaks.

Step 8 — And the ones that exist are falling apart

Panel data, fixed effects, placebo test

Steps 6 and 7 explain the gap through where restrooms are and when they open. There is a third reason, and it comes from a different dataset with none of the weaknesses of the first seven steps.

NYC Parks sends trained inspectors on a schedule to grade every comfort station — the American term for a standalone public toilet building in a park — litter, graffiti, amenities, and structural condition. That is 31,768 inspections, against 3,629 complaints. An inspector who turns up on a rota, not a resident who chooses to call.

Failures more than doubled in two years. Counting the facilities that could not be rated at all — locked, closed, mid-renovation — only 56.8% of 2026 inspections found an open, acceptable restroom, against 71.3% in 2024.

Three things could explain that without any building changing. We tested all three.

Is it a different set of buildings? No. With facility fixed effects — comparing each comfort station against its own history — 2026 still sits +11.4 percentage points above 2024 (SE 2.0pp, p < 0.0000001). The same buildings got worse.

Is it harsher inspectors? No. The same inspector, on the same visit, fills in four scores. Taking only the seven inspectors who rated restrooms in both years, so the grader is held fixed:

A grader who got stricter would mark litter harder too. Litter went down, while structural nearly quintupled. Only the fabric of the buildings did — plumbing, fixtures, partitions.

Is it a rule change? No. A revision to the rating guidance produces a step at one date. This is a steady climb across ten quarters:

2024: .052 · .060 · .103 · .084 → 2025: .146 · .120 · .135 · .169 → 2026: .173 · .190

New inspectors also fail restrooms less often than returning ones (.090 against .210), so the rise is not new staff applying new standards.

Finding 6: the existing stock is deteriorating, and it is structural decay rather than cleaning. This is the only finding in the project measured by a trained observer rather than inferred from complaints.

Why it matters for the decision. Local Law 58 — the 2025 law above — commits New York to planning around 1,150 new restrooms. It does not fund repair. Meanwhile 80 of 736 Parks comfort stations are already under long-term closure — 46 of them for repairs — and across the wider inventory 91 of 1,066 are not operational.

The capital record prices the two paths very differently:

One new build buys roughly two and a half rebuilds of restrooms the city already owns. The city is preparing to double a stock it is not maintaining.

Step 9 — Work out the cheapest fix for each place

Cost analysis, sensitivity testing

We now have 29 neighbourhoods that need attention. Each needs something different, and the first job is to work out what.

Ask what is actually wrong, place by place. The answer rules options in and out.

• New Dorp–Midland Beach — two toilets, and only 19% of the area within a five-minute walk of one. Most of the neighbourhood cannot reach a facility at all, so opening the existing two for longer changes little. The provision is not there.
• East Harlem (North) — seven toilets, good coverage, and still three times the expected complaints. People can reach a toilet. An eighth building is unlikely to be the missing piece, so the problem lies with the seven that exist: their condition or their hours.

Where more than one fix would work, cost decides.

Take a place with toilets in reasonable condition that close at 5pm. Two things would genuinely help — build another one, or keep the existing ones open longer:

Both fix the problem. One is free. So the rule is: take the cheapest option each place's condition allows.

Applied across the 29:

Finding 7: ten of the 29 need nothing built. Their toilets exist, work, and simply close before the hour complaints peak.

What that ordering is worth:

The gap is not really the finding — the cheapest option costing less than the most expensive one is what the words mean. The finding underneath it is that ten of the 29 places already have working toilets that simply close too early.

The thresholds — 35% coverage, "unacceptable" condition — are judgement calls, and the mix of interventions is sensitive to them. We have not run a full sensitivity grid over the thresholds; that is an open item, not a result.

We also tried to prove toilets work. Twice.

Difference-in-differences, natural experiment, placebo test

Both attempts failed, and the second failed in a way worth showing.

Attempt 1 — find toilets that closed, check whether nearby complaints rose. The records turned out to log construction activity, not closures: only 151 of 2,057 flagged entries actually read "Closed". Unusable.

Attempt 2 — about 132 toilets shut every winter and reopen in spring, while 843 stay open year-round. A natural before-and-after with a built-in comparison group.

• It produced a clean result: closure associated with 38% fewer nearby incidents.
• A placebo test, fixed in advance: alcohol summonses have nothing to do with toilets, so they should not move.
• They moved almost as much: 0.699 against the real outcome's 0.624.
• The reason: seasonal toilets sit in parks, and parks empty in winter. Every kind of street offence falls together. We had measured the season, not the toilet.

Finding 8: we publish no effect estimate. A design that fails its own placebo gives no evidence in either direction.

What we recommend

Randomised staggered trial design, power analysis

So: not a construction programme. An experiment — because the number nobody has is the one the city most needs.

• Take the 10 neighbourhoods whose toilets exist and work but close early.
• Extend closing time by four hours, staggered in random order over ten months, so every area is treated eventually and the timing differences do the measuring.
• Those ten contain 44 facilities. At four extra staffed hours a day, about $2.2M a year, or $1.9M for a ten-month trial. No capital spend.
• Run by NYC Parks, which already operates them.

What it could detect:

Complaints are far too sparse to run a trial on — a trial using them could only detect an effect four to five times larger than anything plausible. The trial should measure summonses instead.

That is consistent with what we found earlier: summonses track policing rather than need, which disqualifies them for ranking places, but a before-and-after design compares a place with itself, so a constant level of enforcement cancels out. The alcohol-summons series is available as the same placebo that caught us before.

37% is blunt: it catches a large effect and misses a modest one. It is still the only measurement anyone has attempted, and it costs $1.9M against $34.8M of building.

Alongside it, the one action that needs no experiment. Step 8 found the existing stock deteriorating — 80 comfort stations already under long-term closure, 46 of them for repairs. Restoring a restroom the city already owns costs a median $1.66M against $4.0M to build a new one, and unlike the hours trial it does not depend on an effect nobody has measured: a closed restroom serves nobody, and a reopened one serves whoever is there. Fix before building.

Then spend the capital on new sites, informed by a measured effect rather than an assumed one.

How we know the ranking holds up

Bootstrap resampling, spatial autocorrelation, sensitivity analysis

Five checks:

The one check it does not fully survive is reporting bias — see Complaints measure complainers in the Limits tab.

And one question the table does not answer: how sure are we about a single neighbourhood?

The five checks ask whether the ranking holds together. A budget-holder asks something narrower: is this place really short of restrooms? The model answers it directly — thin counts get pulled towards the citywide pattern, and what survives that pull is what we can be confident about.

Finding 9: six neighbourhoods are individually above 95% confidence, with a seventh a whisker below. Each was re-scored with the model and the prior rebuilt from the other 196, so it could not vouch for itself. All held.

This fixes the thin-counts problem and nothing else — it does not touch what the complaints measure.

What this cannot do

Every limitation, with the arithmetic, is in Limits — the second tab. The three that matter most: the ranking orders places but does not sort them into categories; complaints measure complainers; and the gap between prediction and reality is the one quantity nobody can validate, here or anywhere.