In New York City's midterms, the gap in reported voting between the oldest and youngest adult citizens widens to 19 points, from 12 in presidential years

On the census’s count of adult citizens, in every one of the fifteen elections since 1996 the share of the city’s adult citizens who said they voted was higher than the Board’s ballots per hundred. In midterms the difference ran from 8 to 12 points.

It averaged about 10 in midterms, against about 6 in presidential years, with or without 2020, whose base is the Census Bureau’s experimental release.

The survey’s own count of the city’s adult citizens is itself a sample estimate, controlled to the state’s population rather than the city’s; it ran up to a tenth below the census’s before 2004 and leaves out people in institutions. On that count the midterm difference runs from 6 to 12 points and the presidential average is about 4, and in 2008 the Board’s ballots come out slightly above the share who said they voted.

Inside the margin here means smaller than the reported rate’s own margin of error, with nothing added for the denominators: on the census’s count the difference is inside it in 2004 and 2008, and on the survey’s own count in 2000, 2008 and 2024.

So the difference mixes reporting error with the survey’s sampling error and the gap between the bases, and the survey can’t say whether one group over-reports more than another; the gaps between groups could be wider or narrower than reported, by amounts this check can’t measure.

Leaving them out doesn’t remove it. The rates count a missing answer as not voting, as the Census Bureau’s published rates do (the files reproduce its New York tables exactly on that basis), whether a refusal, a “don’t know” or a person the supplement didn’t reach. In the recent midterms 29% of the under-thirties had no answer recorded, against 19% of people sixty-five and over.

An answer recorded need not be the person’s own, either: the 2024 survey’s documentation says “a single respondent could provide answers for themselves or provide answers for all eligible households members”. Among records with a yes or no recorded, the under-thirties’ midterm rate is 39% and that of people sixty-five and over 58%, a gap as wide as on the full count.

On that basis the gap is also wider in midterms than in presidential years, by about 11 points, and the narrowing since around 2000 holds too, at about 11 points against 7 on the full count. Neither calculation can say what the people without an answer would have said.

These ranges describe uncertainty from sampling, not every source of error, at the 90% level the Census Bureau uses. A range that excludes zero is a difference that chance in the sample is unlikely to explain; one that spans zero means the difference could still be chance.

The midterm age gap is wider than the presidential one by between 2.5 and 13.2 points. It has narrowed since the midterms of 1998 to 2006 by between 1.9 and 12.8 points.

In the recent midterms the gap between the oldest and youngest exceeds the gap between a degree and none by between 0.7 and 10.4 points, the closest of these calls; in presidential years schooling’s gap is the wider, by 1.8 to 10.9 points.

That ranking depends on where the lines are drawn. Age is cut at its extremes, under 30 against 65 and over, and schooling once, a degree against none. Cut age once too, at 30, and the two divide about equally: the difference between the gaps could be anywhere from -2.6 to 5.3 points.

Cut schooling at its extremes instead, no diploma against a graduate degree, and its midterm gap is about 28 points, wider than age’s.

With age cut at 30, its edge over birthplace and renting in the midterms holds, and so does the narrowing since the midterms around 2000; but the widening from presidential years to midterms shrinks to about four points, a range of 0.0 to 8.1, which could still be chance. In presidential years the age gap’s edge over renting’s is not established either; its edge over birthplace’s is.

The survey’s population controls changed between 2002 and 2004, so a pool of its weights gives later elections more weight. With every election weighted equally instead, the widening and the narrowing both hold, the narrowing by 2.7 to 13.5 points.

All rest on margins clustered on the household only, so they are approximate; the Bureau publishes replicate weights for the 2024 supplement, which this analysis doesn’t use. The two closest calls, the under-thirties’ larger drop than the 45-to-64s’ and age’s edge over schooling in the midterms, would span zero if the true margins were about 13% wider; the widening in midterms survives margins 48% wider, and the narrowing 36% wider.

The Census Bureau publishes these rates for New York State, and the files match its published percentages to one decimal place: 49.5% of the state’s adult citizens said they voted in 2018, and 59.2% in 2024, as the Bureau’s tables print them. The city’s figures come from the same records, the five boroughs being identified in every year’s file.

Because one year’s city sample is small: about 1,000 to 1,900 adult citizens, and a few hundred under thirty. Taking the four midterms from 2010 to 2022 together, and the five presidential elections from 2008 to 2024, gives each group enough people for a margin of a few points.

Each combined rate is every weighted yes over every weighted adult citizen across those elections, so a year counts in proportion to its estimated population, not equally; it describes the elections together, not any one of them. The data file gives the weighted numerator and base of every rate it holds, and the Board’s ballots with the census estimates they are divided by.

The files. The Current Population Survey’s November Voting and Registration Supplements for every even-year election from 1996 to 2024: the Census Bureau’s own files from 2010 and NBER’s archive of the same files before. The 1994 file asked the vote question differently, so the series starts in 1996. The fixed-width years are read at the columns each year’s record layout gives; the 1996 layout comes from NBER’s dictionary because the Bureau’s 1996 documentation is a scan. Every column read from the fixed-width years, the household number included, is checked against the layout’s text, and the build stops if one differs or can’t be found; the comma-separated files of 2020 to 2024 are read by their column names.

Who and how. Adult citizens 18 and over in the five boroughs, in the survey’s universe of civilians living in households (no one in an institution, and the armed forces left out), the survey’s final weight. Renters and owners are the adults living in rented or owner-occupied homes, whoever holds the lease or the deed. “Said they voted” follows the Census Bureau: a yes to the vote question over every adult citizen, counting refusals, “don’t know” and no answer as not voting. A degree is a bachelor’s or higher. The age divide compares the youngest and oldest groups, leaving out ages 30 to 64; the other divides split every adult citizen in two. Margins are 90%, clustered on the household in each year; the files read here carry no variance detail, and the Bureau’s replicate weights for the 2024 supplement are not used, so the margins are approximate. As a scale check, applying the Bureau’s variance parameter for New York State, which it provides for state-level totals, to the city’s 2024 estimate for the under-thirties gives a margin of 6.5 points, against 6.9 here. Gaps between two groups use the same households, and comparisons between eras treat them as independent samples.

The check against the count. The Board’s ballots in each election over the census’s estimate of the city’s citizens 18 and over: the 1990 and 2000 censuses and the yearly survey from 2006, with the years between a straight line, an assumption, and 2020’s from the experimental release. 2005 is left out because, in the Bureau’s words, the group quarters population is “not included in the 2005 ACS”.

Code and checks. One script, analysis.py (linked below), produces every figure and chart, and checks the layouts, the Bureau’s published table, each quoted passage against its saved document and each registered comparison against its range; the site’s tests hold the page’s printed figures to the script’s. The dataset cards behind it (cps_voting, nyc_boe_turnout) list the traps.