Mapping the accessibility gap

The MTA is spending $7.1 billion to make about sixty more stations accessible, and by law the choice has to weigh who lives nearby. Earlier analyses worked at the scale of community boards, about 130,000 residents each, which averages away exactly the differences that matter. I worked with census tracts, about 4,000 residents each, all 2,327 of them.

The measure is the accessibility gap: how much further a tract's residents travel to the nearest accessible station than to the nearest station of any kind. Subtracting the nearest station removes ordinary transit coverage, so a neighbourhood with no subway at all has a transit problem, not an accessibility one, and doesn't count against the system. Weighting the gap by the number of residents with a disability gives person-miles, the burden each tract carries.

The gap clusters, and that is the clearest result

Global Moran's I on the gap is +0.70 (p = 0.001): underserved tracts sit next to underserved tracts far more than chance would allow. Local Moran's I finds 161 tracts in significant hot spots, home to 112,151 residents with a disability, and they turn up in every borough: 46 in Brooklyn, 39 in the Bronx, 30 on Staten Island, 24 in Manhattan and 22 in Queens. Much of eastern Queens shows as a cold spot, but its tracts sit on average twice as far from any station as the city's, a transit gap this measure deliberately doesn't count against accessibility.

Not bigger gaps, but more of them

Does the gap fall harder where more people have a disability? Not in the way a correlation would look for. The average gap is flat across tracts ranked by disability rate. What changes is how often a gap exists at all: from about 56% of tracts in the lowest tenth to 69% in the highest (logistic model, odds ratio 1.028 per percentage point, p = 0.001).

Almost nothing survives spatial dependence

An ordinary regression says income, race and density all predict the gap, with very small p-values. But the regression's leftover errors are themselves clustered (Moran's I = +0.67), so those p-values can't be trusted. Refitting as a spatial error model (λ = +0.85) leaves every one of them insignificant. The honest reading is that the gap has strong spatial structure that tract demographics don't explain. Reporting the ordinary regression alone would have produced three confident, wrong conclusions.

Which stations next

For each of the 332 stations that could be upgraded, I recalculated every tract's gap as if that station were accessible, and summed the person-miles removed. Taking the ten best scores on their own claims 52,255 person-miles. Built together, those ten deliver only 28,916, because six of them sit on the Staten Island Railway and the Queens M and R lines and serve the same neighbourhoods.

Choosing one station at a time instead, and re-scoring the rest after each pick, gets 42,443 person-miles from the same ten upgrades: 47% more for the same budget, and a quarter of the citywide burden.

Of the twelve stations the MTA announced in July 2025, one, 63 Dr–Rego Park (highlighted), is also first on this list. The others rank between 14th and 289th on this one measure. The MTA's selection weighs criteria beyond what a distance model can see, so this is a comparison, not a verdict.

How much each choice matters

I re-ran the whole analysis under each alternative choice. Two barely mattered. One mattered enormously: counting a station as accessible only if it has a backup elevator or a ramp. On paper a third of stations are accessible, but 314 of the system's 478 elevators have no backup, so when that one elevator is out, the station isn't accessible that day.

What it adds up to

The gap is real, concentrated, and spatially structured in a way demographics don't explain. It falls on residents with a disability not as bigger distances but as more neighbourhoods with any gap at all. The order of upgrades matters as much as their number: chosen one at a time, the same ten stations remove 47% more of the burden. And the official one-in-three figure flatters the system, because most of it hangs on a single elevator.

Limits

• Distances are straight lines, which understates the gap wherever highways, parks or water break the walking network. Routing on OpenStreetMap is the most valuable next step.
• Disability counts come from a survey, with errors widest where populations are smallest. Screening out unreliable tracts removes 28.5% of them, so results are reported with and without the screen.
• The correlation between disability rate and the gap is weak under every scenario, and its significance depends on that screen. Nothing important rests on it.