Vote Safe Preliminary Voter Vulnerability Report

Racial demographics, registration and turnout across the United States, with six designated focus areas in Ohio, Georgia, Michigan and North Carolina

Prepared by SAMANDA · Save America Movement · 10 August 2026, revised 27 August 2026 · Preliminary — for internal review

1 · Measurement and definitions

1.1 Race categories

The Census Bureau records race and Hispanic origin as two separate questions. The four categories used throughout this report follow the standard convention derived from that structure.

Table sources. U.S. Census Bureau CVAP special tabulation, 2024 vintage — these are that tabulation’s own category definitions, not choices made here.

The four are mutually exclusive but do not sum to 100%: American Indian and Alaska Native, Native Hawaiian and Pacific Islander, Other, and Two-or-More races form the remainder and are carried in the underlying data files. Black and Asian additionally have any-part variants that include people reporting the race in combination with another.

1.2 Two measurement systems

Census enumeration. The P.L. 94-171 redistricting file gives population and voting-age population by race for every census block in the country. The CVAP special tabulation adds citizen voting-age population by race. Neither imputes anything. All precinct-level findings in this report rest on these two sources.

Voter file. Eight states ask race or ethnicity on the voter registration form — Alabama, Florida, Georgia, Louisiana, North Carolina, Pennsylvania, South Carolina and Tennessee. Pennsylvania asks but the field is effectively empty, holding 3,247 self-reported records against 711,030 assigned by vendor; seven states therefore yield usable observed race. In the remaining 43 states the record carries no race field at all, and L2 — a commercial data vendor — imputes race from surname, given name and the racial composition of the voter's census block.

2 · Demographic distribution

The national citizen voting-age population is 237.7 million: 155.7 million white, 33.3 million Latino, 29.1 million Black and 11.5 million Asian. Its geographic distribution is highly uneven, and the unevenness is itself a finding — it determines whether an intervention organized by county reaches most of a group or only a fraction of it.

3 · Registration

Registration by race can only be measured from a voter file, and therefore carries the limitation set out in §1.2. Nationally the voter file records 215.7 million registrants.

Table sources. Registration counts: L2 2024 general-election statistics, aggregated to 2020 census blocks by the Redistricting Data Hub. Citizen denominators: U.S. Census Bureau CVAP special tabulation, 2024 vintage. The rates are computed in this report; L2 publishes the counts.

Only one jurisdiction in the reference collection publishes registration by observed race below the county level: Louisiana, whose legislature releases a precinct-level file giving registered voters by race and by race × party for 3,539 precincts. Six other states that collect race publish comparable statistics that are not yet held.

4 · Turnout

4.1 By district and county

4.2 By precinct composition

4.3 Normalized comparison

Raw turnout bands mix two effects: differences between racial groups and differences between states. Scoring every precinct against its own state's white-precinct benchmark removes the second, so that a precinct in Minnesota and one in Kentucky are compared on the same scale.

This comparison is shown for every state at once, and one state at a time. An earlier version of this figure plotted only the pooled curve. That curve answers the question it was built for — whether the participation gradient is merely an artifact of which states have large non-white populations, and it is not — but on its own it reports an average shape that no individual state actually has. Across the 40 states that can be built, the gradient is present everywhere, but its steepness ranges from 63 index points in AK to 19 in WV, a spread of roughly 3.3 to one around a median of 38. The panels below keep the pooled comparison and add that variation.

4.4 By state

4.5 Turnout by race from the voter file, and the one group it is observed for

Two outside sources appear in this section, and neither is a census of who voted. L2 is a commercial voter-file vendor: it buys each state’s public registration and vote-history records, standardises them, and sells the result. Where a state records a registrant’s race on the registration form, L2 carries that entry through. Where a state does not, L2 assigns race with a model of its own. Its shipped documentation says only that “L2 ethnicity categories use modeling techniques to infer an individual’s ethnicity” and that “self-reported fields come directly from voter-provided descriptions”. It does not publish the method, and it is not the BISG procedure described in section 9 — a name this report does not attach to it, because nothing in L2’s documentation supports it. It is proprietary and name-based, and it can only be checked from outside against a state that records race, which is what this section does. CPS is the U.S. Census Bureau’s Current Population Survey, whose November supplement asks a sample of households whether they voted; it is a survey of about sixty thousand households, and people over-report voting in it.

Georgia and North Carolina record race on the registration form. Michigan and Ohio do not. That single difference decides what this section can say, and it is measured below rather than assumed. Black is the only group for which any state’s own record exists in this data. The L2 product carries one self-reported race field and it is Black; there is no self-reported Latino, Asian or white field for any state, so every Latino figure below — in Georgia and North Carolina as much as in Michigan and Ohio — is L2’s model, not a count.

One qualification, because it bears on what comes next. That holds for THIS product — L2’s 2024 turnout statistics, whose only self-reported field is African-American. A different L2 product, the block-aggregated voter file, does carry the state’s own entry for every category (white, Hispanic, East Asian, Korean, Native American, other, not stated), and it is held for 35 of 51 units including all four focus states. It cannot be used here because it reports REGISTRANTS as of 2022 rather than who voted in 2024 — but it is the obvious next check on section 4.6, and section 9 lists it.

Table sources. Race, registration and vote history: L2 2024 general-election statistics, aggregated to 2020 census blocks by the Redistricting Data Hub, held in SAM Reference-Data at demographics/other/l2/{ST}/{ST}_l2_2024_gen_stats_2020block.zip. Citizen denominators: U.S. Census Bureau CVAP special tabulation, 2024 vintage. Survey column: U.S. Census Bureau Current Population Survey, November 2024 voting and registration supplement. The three ratio columns are computed in this report from those files; L2 publishes the counts, not the rates.

Read the fourth column first. It divides the registrants L2 identifies as a group by the citizens the Census counts in it — how much of that population the file sees at all. Where the state records race it sits near 1.0. In Ohio, where it does not, it is 0.70 for Black citizens: L2’s model finds about 70% of them. Georgia’s Black figure exceeds 1.0, which is not a rounding artefact but the plain meaning of the number: the model assigns more Black registrants than Georgia has Black citizens. A rate built on a denominator like that cannot be repaired by choosing a different numerator.

The fifth and sixth columns divide the same voters by different denominators. The fifth uses the Census citizen count, so it answers “what share of this group’s citizens voted” — and runs low wherever L2 fails to identify people, because the numerator misses them and the denominator does not. The sixth uses L2’s own registrant count, which is internally consistent but answers a narrower question: turnout among the people L2 found, not among the group.

So this section reports turnout by race only for Black citizens in Georgia and North Carolina. Everything else in the table is shown to make the difference visible, not because it is a measurement. For the modeled rows the level is not recoverable from this source, and no correction is applied to make it appear so: any single factor that closes the gap to the survey column is defined by that gap and cannot be checked against anything. The survey column is a comparator and never a target — it is self-reported, and self-reports overstate voting.

Recovering group turnout where the voter file cannot is a different method entirely, and it uses no voter file at all. That is section 4.6.

4.6 Estimated turnout by race, and the seven states it comes from

Section 4.5 reports what a voter file can support. This section reports an estimate made a different way. Ecological inference reads only two things for each precinct — how many people voted, and how many citizens of each racial group live there — and works out which combinations of group turnout rates are consistent with both across thousands of precincts at once. No voter file, commercial or official, enters the calculation.

Three different things are being compared in this report, and they fail in different ways. L2’s model, described in section 4.5, guesses an individual’s race from their name and address and then counts the guesses; the published form of that approach is BISG, Bayesian Improved Surname Geocoding. It is wrong when a name is not diagnostic — a Black registrant with a common surname in a mixed neighborhood — and its errors are correlated with exactly the places this report is about. The CPS survey asks people directly, so it needs no model at all, but it samples about sixty thousand households nationally, reaches nowhere near precinct level, and is self-reported by people who over-state voting. King’s ecological inference, used here, guesses nobody’s race. It takes the Census count of who lives in a precinct as given and asks what group turnout rates would be arithmetically consistent with the ballots that precinct actually counted. Its failure mode is neither a bad name-guess nor a small sample: it is a precinct whose composition is too lopsided for its total to constrain the answer, which is the limitation this section returns to below.

What “behind” is measured against. Every shortfall in this section and on the maps is measured against that state’s own white-precinct benchmark — the turnout of precincts in the same state that are at least 75% white by citizen voting-age population, weighted by their citizen counts. It is deliberately a within-state comparison: turnout differs between states for reasons that have nothing to do with race — competitive races, vote-by-mail rules, registration deadlines — and comparing Ohio to Georgia would fold all of that into a figure meant to be about participation by race. A state whose white precincts turn out at 70% sets a 70% bar for itself. The benchmark is computed only where the 75%-white band holds at least twenty thousand citizens, so a state without a substantial white-precinct population sets no bar and is left out rather than compared to a shaky one.

The model was fitted on seven states: Alabama, Florida, Georgia, Louisiana, North Carolina, South Carolina and Tennessee. Those seven were chosen because each records race on the voter registration form, which means the estimate has a known answer to be wrong against. Georgia and North Carolina are among them. Michigan and Ohio are not. Their figures are produced by applying the seven-state result to states the model was never fitted on, and neither state records race, so there is no independent answer to check them against. Everything below for Detroit, Cleveland and Columbus is an extrapolation, and is marked as one in the table and in each figure.

The maps show votes at stake: a precinct’s non-white citizens multiplied by how far their estimated turnout falls below the turnout of white precincts in the same state. Red is that shortfall, deeper as it grows; blue is the opposite, votes cast above the benchmark. It is the quantity section 4.5 could not compute, because it had no turnout rate for the group by itself.

Table sources. Estimated turnout by group: King’s ecological inference, held in SAM Reference-Data at elections/derived/rxc_turnout_by_race/, both configurations distinguished by that file’s model column. Citizen counts and benchmarks: Citizen voting-age population by race: U.S. Census Bureau CVAP special tabulation, 2024 vintage (2020 census blocks) · Votes: certified 2024 general election precinct returns, allocated to 2020 census blocks. The rates are the model’s; the aggregation, the shortfall and the votes at stake are computed in this report.

The six areas total 514,656 votes at stake across 1,966 majority-non-white precincts. 309,824 of that total is in Georgia and North Carolina, where the model was fitted; 204,832 is in Michigan and Ohio and is extrapolated. A precinct with no estimate is drawn neutral and counts in neither column, which is why the last two columns can fall short of the precinct count.

The counties column is the scope, and it is wider than the area name. Each area is a core of one to three counties — the ones the name lists — drawn together with the counties adjacent to them, because a metropolitan electorate does not stop at a county line and a map that cut it there would report a smaller shortfall than exists. Detroit is the exception at one county, for the reason given below. Section 6 reports the same estimate for the core counties ALONE, county by county, so its totals are smaller than these by construction and the two are not in conflict.

A second estimator, and what the two disagree about. The same method was run again a different way: instead of fitting seven states and carrying the result outward, it was fitted directly on all thirty-nine. Both are reported here, because the distance between them measures how much of an estimate comes from the election rather than from the model. (A third configuration, fitted on a twenty-thousand-precinct sample of the same thirty-nine states, is published with them; it behaves like the full refit — 13.7 points of error for non-white citizens against the survey where the full refit gives 14.0 — and is not shown separately below.)

Table sources. Estimated turnout by group: King’s ecological inference, held in SAM Reference-Data at elections/derived/rxc_turnout_by_race/, both configurations distinguished by that file’s model column. Citizen counts and benchmarks: Citizen voting-age population by race: U.S. Census Bureau CVAP special tabulation, 2024 vintage (2020 census blocks) · Votes: certified 2024 general election precinct returns, allocated to 2020 census blocks. The rates are the model’s; the aggregation, the shortfall and the votes at stake are computed in this report.

The last column explains the one before it. A precinct’s own arithmetic narrows each group’s possible turnout to a range — tight when the group is most of the precinct, wide when they are a small slice. A low number there means the precincts themselves nearly settle the answer. Detroit’s are the most informative of the six and the two estimators land within three and a half points of each other; Columbus’s say least and the two differ by ten. Where the arithmetic runs out the model fills the gap, and that is exactly where two models disagree.

Which figure this report uses, and why. The totals above and the maps below are the FIRST estimator; the second is shown beside them rather than averaged in. The first is closer to both independent comparators on the quantity reported here: against the Census survey across thirty-nine states its error for non-white citizens is 5.8 points against the second’s 14.0, and against the registration form in the seven states that record race, 6.1 against 7.9. The second also shows a defect the first does not — its estimate tracks how little a precinct could say, at a correlation of −0.48 against the first’s −0.02, which is the signature of a model answering where the data did not. Read 514,656 as the figure this report stands behind and 678,553 as the same question answered by a model with a known bias — not as a confidence interval.

Why these six areas survive a finding that sank the national numbers. Run across the whole country, the second estimator’s figure for non-white citizens is unusable: biased low by about fourteen points, because nationally half of all precincts hold too few non-white citizens for their returns to say anything about non-white turnout. This report never uses those precincts. It uses precincts where non-white citizens are the MAJORITY, and there the arithmetic is informative — across the four states a typical range of 0.32, against 0.81 for every precinct in those same states, with 4% uninformative against 70%. The report works in the part of the data that can answer the question.

What these figures will and will not bear. Both estimators reproduce much less variation between precincts than the recorded rates in the seven states show, so the ordering — which precincts and which blocs are furthest behind — is firmer than the size of any single gap. The totals are the scale of a shortfall, not a count of identifiable non-voters. Section 8 states the limitation in full.

Detroit is drawn as Michigan’s Wayne County alone. The surrounding counties are named in the adjacency used elsewhere in this section, but their 2024 precinct boundaries do not resolve completely against the turnout table, and a county drawn with some of its precincts missing would read as a county with fewer voters rather than as an incomplete map.

5 · Votes at stake

For each precinct, the shortfall is defined as the group's citizen voting-age population multiplied by the difference between the state's white-precinct benchmark turnout and that precinct's actual turnout, floored at zero. It is a size-adjusted measure of the participation gap. It is not a forecast, and it is not a mobilisation target: it states how many additional votes would have been cast had the precinct turned out at the rate of that state's white precincts.

5.1 National distribution

5.2 Leading regions by state

Table sources. U.S. Census Bureau CVAP special tabulation, 2024 vintage, aggregated to county and state in this report. No turnout and no model enter this table.

The final column is operationally the most consequential. In Michigan and Arizona a county-level approach reaches three-quarters or more of the state's total; in Georgia, North Carolina and Virginia it reaches under half, and the remainder is distributed across many smaller jurisdictions.

5.3 Geographic concentration

Two measures are reported for each state: the index of dissimilarity against the white population computed across precincts, where 0 is an even spread and 1 complete separation; and the share of the group's statewide CVAP resident in its five largest counties.

Table sources. U.S. Census Bureau CVAP special tabulation, 2024 vintage. The dissimilarity indices and top-five shares are computed in this report from those counts.

Michigan and Ohio hold the most segregated Black populations among the focus states. Georgia and North Carolina are the least concentrated of those shown, and Latino populations are less segregated than Black populations in every state listed.

6 · The six Vote Safe focus areas

Table sources. Votes: certified 2024 general election precinct returns, allocated to 2020 census blocks · Citizen voting-age population by race: U.S. Census Bureau CVAP special tabulation, 2024 vintage (2020 census blocks) · Block-to-precinct assignment: SAM Reference-Data block spine v1.0, from the Census 2020 Block Assignment File. Every column is counted or divided from those three files in this report; nothing in this table is modelled or estimated.

Shares are of citizen voting-age population and are the Census race lines: White, Black and Asian are non-Hispanic alone, Latino is Hispanic of any race. They do not sum to 100% — American Indian/Alaska Native, Native Hawaiian/Pacific Islander, other and two-or-more races are counted in non-white but have no column of their own here.

The same six areas resolved to their eight constituent counties. Two of the areas combine more than one county, and the counties inside them are not alike: in Metro Atlanta, DeKalb is 67% non-white against Fulton's 58%, and Gwinnett turns out 1.9 points above Fulton. Figures sum exactly to the area totals above.

Table sources. Votes: certified 2024 general election precinct returns, allocated to 2020 census blocks · Citizen voting-age population by race: U.S. Census Bureau CVAP special tabulation, 2024 vintage (2020 census blocks) · Block-to-precinct assignment: SAM Reference-Data block spine v1.0, from the Census 2020 Block Assignment File. Every column but one is counted or divided from those three files in this report. The exception is Est. non-white turnout, which is modeled; its source is the one named under the next table.

Turnout here is the precinct's turnout aggregated over the county — votes cast divided by citizen voting-age population, all races together. It is not turnout for any one racial group: precinct returns are a single total, and no by-race vote count exists in any source used here. Est. non-white turnout is the estimate section 4.6 describes, under the seven-state fit this report leads with; an asterisk marks a state that fit was not fitted on. It is measured over the county's majority-non-white precincts on 2024 boundaries, so it rests on a smaller and differently drawn set of units than the columns to its left, and the two are not arithmetically reconcilable. Cluster counts are area-level and are not shown by county, because contiguous blocs cross county lines — metro Atlanta's largest spans Fulton and DeKalb; a per-county figure is carried separately in FOCUS_counties_2024.csv.

The two estimator configurations, per county

Section 4.6 reports both configurations for each area as a whole. The same two, per county, are below. Basis is the seven-state fit's: three of the eight counties sit in a state it was not fitted on, and are extrapolations there. Both thirty-nine-state configurations were fitted on every state in the file, so nothing is an extrapolation under the refit — which is a difference in what the two numbers ARE, not only in what they say. Gap is the distance between the two configurations, and is the honest width of this estimate: where the configurations agree the precincts settled the answer, and where they diverge the fitted curve did. Votes at stake is the seven-state fit's, on the report's formula — each precinct's non-white citizens multiplied by how far their estimated turnout falls below the state benchmark, summed per precinct so a precinct at or above the benchmark contributes nothing rather than cancelling a deficit elsewhere. It is a narrower scope than section 4.6's area totals, which also draw the counties adjacent to these eight.

Table sources. Estimated turnout by group: King’s ecological inference, held in SAM Reference-Data at elections/derived/rxc_turnout_by_race/, both configurations distinguished by that file’s model column. Citizen counts and benchmarks: Citizen voting-age population by race: U.S. Census Bureau CVAP special tabulation, 2024 vintage (2020 census blocks) · Votes: certified 2024 general election precinct returns, allocated to 2020 census blocks. The rates are the model’s; the aggregation, the shortfall and the votes at stake are computed in this report.

A precinct is counted in scope where a majority of its citizen voting-age population is not white and its turnout fell below the state's white-precinct benchmark.

6.1 Precinct turnout

6.2 Adjacency clustering

In-scope precincts were grouped into connected components by polygon contiguity, so that each cluster is a spatially continuous area rather than a ranked list of separated units.

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Two contiguous blocs account for 200,260 votes at stake between them: a 233-precinct area of metro Atlanta and a 465-precinct area of Detroit, each spatially continuous. Cleveland adds a third bloc of comparable reach — 303 contiguous precincts holding 247,870 citizens of voting age, 81.1% of them not white, which turned out at 45.8% against Ohio's 67.9% benchmark and carries 45,219 votes at stake. Franklin County produces the most extreme rates rather than the largest volume — its two principal clusters recorded 47.3% and 44.4% turnout against the same benchmark — deficits of 20.6 and 23.5 points.

Fragmentation differs materially between areas and bears on operational design. Detroit concentrates 92% of its in-scope shortfall in a single cluster and Cleveland 82% of its own in 8 clusters; Charlotte and Durham are the most consolidated at 6 clusters between them; Franklin County is the most fragmented at 18, with the two largest holding 79%.

6.3 Clusters by group

The clusters above are blocs of precincts that are majority non-white and below benchmark. That answers where the aggregate shortfall sits, but not whose it is: a precinct that is 45% Black and 40% Latino falls in the same bloc as one that is 85% Latino. The two tables below re-run the same contiguity rule over precincts where a named group is the largest by citizen voting-age population and turnout is below the state benchmark, so each bloc is about one group.

67 Black-plurality clusters hold 325,391 votes at stake; 11 Latino-plurality clusters hold 12,448. The asymmetry is the finding, not an artifact: in these six areas Latino citizens are far more often a substantial minority of a precinct than its largest group, so they appear in the non-white clusters above without forming many blocs of their own.

How tightly packed a bloc is, and how much sits in it. Every cluster table now carries the land area of the bloc in square miles, the mean distance between its precincts’ center points, and votes at stake per square mile. The two distances answer different questions: a programme can work a bloc whose precincts average two miles apart even if it runs thirty miles end to end, and cannot work one that averages thirty. Area alone does not separate those, because a long ribbon and a compact disc can cover the same ground. The most constrained blocs of five precincts or more are Columbus cluster 5 (5 precincts, 2 sq mi, centres 1.4 mi apart, 955 votes at stake), Columbus cluster 6 (9 precincts, 3 sq mi, centres 1.7 mi apart, 843 votes at stake), Cleveland cluster 2 (40 precincts, 7 sq mi, centres 1.9 mi apart, 8,331 votes at stake). Smaller blocs are measured but not ranked: a one-precinct cluster has a mean center distance of zero and would otherwise top every list.

Black-plurality clusters, largest 8 by votes at stake

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino-plurality clusters, largest 8 by votes at stake

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

6.4 Where the clusters are

6.5 Largest precincts by group

The twenty precincts in each area holding the most citizens of each group. Ranked by population, not by share — a share ranking selects small precincts, and the question here is where the most eligible voters of a group are. The full table, all 6 areas by five groups, is FOCUS_top20_by_area_2024.csv.

Black — Columbus, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Black — Cleveland, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Black — Metro Atlanta, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Black — Detroit, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Black — Charlotte, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Black — Durham, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Columbus, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Cleveland, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Metro Atlanta, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Detroit, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Charlotte, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

Latino — Durham, largest 10 precincts

Table sources. Certified 2024 general election precinct returns · Census CVAP special tabulation, 2024 vintage · SAM Reference-Data block spine v1.0 · Census TIGER/Line boundaries. Computed in this report from those files.

7 · Methods

Population and eligibility

Population and voting-age population by race were taken from the Census P.L. 94-171 redistricting file at the 2020 census block, for 51 states and the District of Columbia. Citizen voting-age population by race was taken from the Census CVAP special tabulation, 2024 vintage, disaggregated to the same blocks by the Redistricting Data Hub, using the non-Hispanic-alone race lines with their published margins of error. State and county race distributions were additionally drawn from American Community Survey table B03002, 2020–2024.

Geographic assignment

Every figure in this report is built by counting up from individual census blocks — the smallest unit the Census publishes — and adding them into precincts. Because 2020 voting districts are constructed from whole blocks, no block is ever split between two precincts, so this is exact addition rather than estimation. Both population and votes are aggregated the same way, which is what makes a turnout rate meaningful at precinct level.

Blocks were assigned to precincts using the voting-district component of the Census 2020 Block Assignment File. Blocks were assigned to congressional districts using the Census 119th Congress Block Equivalency File — 8,174,955 blocks across 52 units, with approximately 200 unmatched nationally. Derived crosswalks held in the reference collection were not used, as they carry the district map in effect in January 2021.

Turnout

Votes were taken from 2024 general election precinct returns allocated to 2020 blocks, summing all presidential contest lines. Turnout is expressed as votes divided by citizen voting-age population, a citizen-based rate comparable to Current Population Survey and voting-eligible-population rates. Precincts with fewer than 50 citizens of voting age, or with implied turnout above 110%, were excluded as allocation artefacts.

Benchmark and shortfall

Each state's benchmark is the CVAP-weighted turnout of its precincts that are at least 75% white by citizen voting-age population, computed only where that band holds at least 20,000 CVAP. In plain terms it answers: in this state, what did turnout look like in the places that are overwhelmingly white? Every other precinct in that state is then measured against it, so that a precinct is never compared with a different state's political weather. The benchmarks are Michigan 77.9%, Georgia 75.4%, North Carolina 78.2% and Ohio 67.9%; all 31 are in FOCUS_state_benchmarks_2024.csv. The 20,000-CVAP floor is a real gate: Hawaii's band holds only 3,946 CVAP and Alabama's none, so neither state receives a benchmark.

Votes at stake is the sum over precincts of group CVAP × (benchmark − precinct turnout), floored at zero — the number of additional votes a precinct would have produced had it turned out at its state's white-precinct rate. It is an arithmetic description of a gap, not a forecast and not a mobilisation target, and it is floored at zero so that precincts above their benchmark do not offset those below. Area totals in section 6 sum this over in-scope precincts only.

In scope marks a precinct where a majority of citizen voting-age population is not white and turnout fell below the state benchmark. Both conditions must hold; either alone is not sufficient.

The normalized index (Figure 6)

Each precinct's index is its turnout divided by its own state's benchmark, multiplied by 100, so 100 means it matched that state's heavily-white precincts. Precincts are sorted by the non-white share of their CVAP and divided into ten equal-count groups — each decile holds the same number of precincts, not the same population. The value plotted for a decile is the CVAP-weighted mean index of its precincts, equivalently the decile's total votes divided by its total CVAP, divided by the benchmark. Weighting by CVAP means a decile is not swung by very small precincts. Precincts excluded as allocation artifacts are excluded here too.

Because each state is indexed to its own benchmark before any grouping, differences in overall state turnout are removed and what remains is the gradient within states. Pooling states into a single curve is therefore legitimate as an answer to “is the gradient just a state effect?” but it averages away real differences in steepness, which is why this revision reports the states separately.

Figure 6 is built on the same precinct universe as the rest of this revision. That was not true of the 10 August draft, which built it on the 2024 election precinct because voting-district boundaries were then held for only 18 states and the block-to-district assignment could not be derived nationally. It now can: the collection publishes a block-level spine carrying each block's voting district, and every table and figure rebuilt here is counted up from it. The earlier substitution, and the caveat that went with it, are retired.

Rebuilding on the spine moved some counts, and that is the point of doing it rather than a cost of doing it — the assignment underneath is more reliable, so the figures resting on it are more nearly right. Two Wayne and Cuyahoga units disappeared from the focus tables because they were Census “ZZZZZZ” districts — open water and unorganised territory — that the previous point-in-polygon method had placed blocks into. Both had already been excluded from every computation, so no benchmark, in-scope count or votes-at-stake figure changed. At national scale the gradient is unchanged: the pooled index falls {nat['pooled'][0]['index']:.1f} to {nat['pooled'][-1]['index']:.1f} across {len(nat['per_state'])} states, against 104.4 to 66.8 across 33 in the 10 August draft.

Concentration

The index of dissimilarity is computed across precincts within each state as ½ Σ |b_i/B − w_i/W|, where b and w are the group and white citizen voting-age populations of precinct i and B and W their state totals.

Race categories

The Census records race and Hispanic origin as two separate questions. Throughout this report White, Black and Asian mean that race alone and not Hispanic, while Latino means Hispanic or Latino of any race. The four are mutually exclusive but do not sum to the total: American Indian and Alaska Native, Native Hawaiian and Pacific Islander, Other, and two-or-more races make up the remainder and are carried in the underlying data files. Where this report says non-white it means total CVAP minus white CVAP, which therefore includes those remainder groups and is slightly larger than Black plus Latino plus Asian.

Adjacency

Precinct polygons were taken from TIGER/Line 2020 voting districts, projected to the appropriate UTM zone, and grouped into connected components by a union-find over intersecting geometries with a one-meter tolerance to absorb digitising slivers.

8 · Limitations

1. Ecological inference. Precinct-level results describe precincts, not individuals. A turnout rate of 52% in precincts that are at least 75% Black does not establish that 52% of Black citizens voted: some residents of those precincts are not Black, and most Black citizens live elsewhere. Recovering individual-level rates requires formal ecological inference, for which the precinct file is the standard input.
2. Imputed race. Registration figures outside the seven observed-race states rest on a commercial model that under-identifies Black registrants by roughly half. Turnout rates from that source remain usable because the error affects numerator and denominator alike; counts do not.
3. Coverage. The block spine carries 51 units — the fifty states and the District of Columbia. Of these, 41 build a unit table and 40 enter the normalized index of Figure 6. Ten are excluded, each for a reason the build measures rather than assumes: Maine and Oregon carry no spine column complete enough to key on, the most complete covering 38.5% of Maine's blocks and 0.0% of Oregon's blocks against a 99% floor; the District of Columbia places only 0.0% of its spine citizen population inside a 2020 voting district, against a 98% floor; Arkansas, Indiana, New Jersey, Oklahoma and Pennsylvania have no block-allocated 2024 returns in the collection, and being keyed to the 2020 voting district their precinct returns cannot be joined without a precinct-to-unit crosswalk; and Mississippi and South Dakota join too few of their votes to the spine's precinct identifier, 0.0% and 83.8% against a 98% floor. Hawaii builds a unit table but is held out of the normalized index, whose white-turnout benchmark band holds only 3,946 citizens there. The 10 August draft named California, Hawaii and Oregon as unresolvable; on this build only Oregon is, and California resolves in full — 22,048 units, 16,218 of them in the index.
4. Unknown-race registrants. In the voter file the unknown category is 8–11% of registrants, turns out above the local mean, and is disproportionately white. Subgroup shares should be reported both including and excluding it.
5. Single cycle. All findings describe 2024 alone and cannot distinguish a durable gap from a cycle-specific one.

9 · Proposed additional analysis

The following extend the present work using data already held in the reference collection, ordered by expected value relative to effort.

1. Ecological inference per state, with covariates — and a Bayesian hierarchical model beside it. Section 4.6 does what this item used to propose, twice over, and the second run identified the specific repair. Both fits pool every state into one national curve, so where a precinct's own arithmetic cannot settle the answer that single curve fills the gap — which is why the two estimators diverge most exactly where precincts say least. Two changes follow. Fit each state separately, so a state's fallback is its own electorate rather than the country's. And give the fallback something real to stand on: prior-cycle turnout, age structure, population density, language, vehicle access and tenure are all held at block-group or tract level in the collection already. A Bayesian hierarchical specification is the natural vehicle — it can pool partially across states rather than all-or-nothing, carry those covariates as predictors, and use the block-adjacency files held for all fifty states as a spatial prior so neighbouring precincts inform each other. It also reports an honest posterior, which the present method does not: section 4.6 explains why the interval King's EI states is not usable.
2. North Carolina's own voter file — acquired, and the ground truth this work has lacked. This item previously proposed the acquisition. It is done: the statewide registration file, the voter history file, and 72 registration snapshots spanning November 2005 to March 2026 (65.2 GiB) are held in the reference collection. Its registration record carries race and ethnicity as reported by the registrant, so North Carolina turnout by race becomes a direct count rather than an estimate — and, more valuable, a recorded answer against which the estimator in section 4.6 can be scored at PRECINCT level rather than only against a statewide survey. The layout carries both a precinct and a voting-district identifier natively, so no crosswalk stands between the file and the maps in this report. The snapshot archive also supplies the trajectory item below for one state without waiting on anything.
3. Individual state data beyond North Carolina. Seven states record race on the registration form itself — Alabama, Florida, Georgia, Louisiana, North Carolina, South Carolina and Tennessee — and those seven are the only places any estimate of turnout by race can be checked against a recorded answer. North Carolina is now held. The same intake path applies to the other six and to states that publish a file without a race field (Michigan, Ohio and Washington among them), which are still worth holding for vote history, address and registration-date detail that supports the covariates above. Each acquisition turns on that state's own published terms, which differ and must be read per state rather than assumed from North Carolina's: that intake recorded the licence position, the personally-identifying fields the publisher withholds, and the file depth as separate decisions, and the same three questions have to be answered again for each state. Where a state charges for its file or restricts redistribution, that is a procurement decision rather than a technical one, and this report does not assume the answer.
4. Trajectory, 2016 to 2024. Crosswalk builds are underway. Precinct returns exist for 2016, 2018, 2020, 2022 and 2024, and eligibility by race exists annually from 2015. A four-cycle series separates durable gaps from 2024-specific ones and identifies areas that are deteriorating rather than merely low. The obstacle is not the returns but the geography: precinct boundaries move between cycles, so a precinct in 2016 is frequently not the same place as the precinct with that name in 2024, and a series built without reconciling them measures boundary change as if it were behavioural change. The crosswalk library that resolves this is being built now. North Carolina's snapshot archive is the exception that needs none of it — the same registrants are followed through 72 dated files.
5. Other ways to assign race where no state records it. L2 does not publish how its model works — its documentation says only that it uses “modeling techniques to infer” ethnicity — so it cannot be audited, only scored against a state that records race. The established published alternative is BISG — Bayesian Improved Surname Geocoding — which combines the Census surname list with the racial composition of a registrant's block group; BISFG adds the first name, which materially improves classification for Black registrants in particular, the group this report most depends on. Both are reproducible from data already held: the Census surname file, block-group composition, and a geocoded address. Their advantage over a vendor's model is not accuracy but auditability — the inputs and the arithmetic can be published, and the error can be measured directly against North Carolina's recorded race. Before any of that, there is a check available today: L2's own block-aggregated voter file carries the STATE-REPORTED entry for every category (eth3_*) for 35 of 51 units, including all four focus states, and the catalogue records the model running 12.8% above the state's own Black count and 49% above its Hispanic count in Georgia, with 9.35% of registrants stating no race at all. Scoring the model against that file measures the error directly rather than inferring it. Machine-learning and deep-learning classifiers trained on recorded ground truth are a further step, and a genuinely promising one now that a labeled corpus of several million registrants exists in the collection; a model trained on North Carolina and tested on Georgia would measure how far such a classifier transports, which is the question that decides whether it may be used in Michigan or Ohio at all. The discipline that applies to all three is the one section 4.6 already states: a method may only be used where its error against a recorded answer has been measured.
6. Boundary layers that would sharpen this. The collection holds considerably more geography than this report draws on, and several layers bear directly on the open questions. County subdivisions (950 files) are the township and municipality layer — the unit Michigan actually administers elections in, and the one whose cross-county municipalities account for the precincts this report could not match. Census tracts and block groups (945 files each) are the grain at which the socioeconomic covariates above are published, and so are the natural carrier for a covariate-informed model. PUMAs (698) link to the Census microdata sample, the only public source of individual-level demographics that could train or validate a classifier without a voter file. School districts (1,005 unified, 459 elementary, 356 secondary) and state legislative districts (863 upper, 828 lower) are administrative overlays that turn a precinct finding into a finding about a body that can act on it. Roads and water (3,231 and 3,235 files) matter for the polling-place item below: distance along a road network is a very different measure of access from distance across a river. And the block adjacency files held for all fifty states are what a spatial prior in the hierarchical model would be built from.
7. Administrative burden. Election Administration and Voting Survey county files for 2014 through 2024 carry provisional ballots, rejected mail ballots, poll-worker counts and wait times. Joined to precinct composition these test differential burden — a mechanism — rather than differential outcome.
8. Polling-place access. Polling-place locations are held for the focus states. Distance from precinct centroid to nearest polling place, and change in that distance across cycles, is directly computable and is among the most legible measures of access.
9. Redistricting counterfactuals. The collection holds North Carolina's full congressional plan history, including the plan invalidated in 2021 and those adopted in 2023 and 2025. Recomputing Black citizen voting-age population by district across those plans quantifies the effect of each map.
10. Prison-adjusted denominators. Prisoner-reallocation files exist for all 50 states. Prison gerrymandering relocates population between districts; the counterfactual files measure the shift rather than assert it.
11. Forward projection. Voting-age population projections to 2035 are held at block and county level with race detail. Running the votes-at-stake calculation forward indicates where the arithmetic moves by 2028 and 2032.
12. Groups beyond the four. American Indian and Alaska Native populations are invisible in the present framing yet are geographically concentrated enough to be tractable; the block file carries them, along with Native Hawaiian and Pacific Islander and Two-or-More.
13. Language access. American Community Survey language tables support identification of jurisdictions near Section 203 coverage thresholds, on the same geographies used here.
14. Socioeconomic covariates. Approximately one hundred ACS table families carry the full race iteration — income, educational attainment, employment, tenure, insurance and disability. These move the analysis from whether participation differs to what explains it, without new acquisition.

10 · Data sources

1. United States Census Bureau. 2020 Census Redistricting Data (P.L. 94-171) Summary File. Block-level population and voting-age population by race and Hispanic origin.
2. United States Census Bureau. Citizen Voting Age Population (CVAP) Special Tabulation, 2024 vintage; block-level disaggregation by the Redistricting Data Hub.
3. United States Census Bureau. American Community Survey five-year estimates 2020–2024, tables B03002 (Hispanic or Latino origin by race) and B04006 (people reporting ancestry).
4. United States Census Bureau. 2020 Census Block Assignment Files, voting-district component.
5. SAM Reference-Data. Block spine, v1.0, 2026 — one row per 2020 census block carrying its voting district, 2024 election precinct, citizen voting-age population and congressional district. Derived from the Census sources above; it is what assigns every block in this report to a precinct.
6. United States Census Bureau, Redistricting Data Office. 119th Congress Block Equivalency Files, 2025.
7. United States Census Bureau. TIGER/Line Shapefiles 2020, voting districts; cartographic boundary files, 2023 and 2024.
8. Redistricting Data Hub and Voting and Election Science Team. 2024 general election precinct returns allocated to 2020 census blocks.
9. L2 Inc., via the Redistricting Data Hub. Voter file aggregates by county, voting district and census block, 2020, 2022 and 2024.
10. United States Census Bureau. Current Population Survey, November Voting and Registration Supplement, 2008–2024.
11. Louisiana Legislature. 2020 Registered Voters Equivalency File, voting-district level, retrieved by the Redistricting Data Hub.