Comparing Precincts Across Two
Elections
Precinct names and precinct lines drift
independently, so a precinct that swung may be a precinct that
moved
Democracy’s Data
Last updated September 15, 2026 at 4:42p EDT
Every account of an election contains sentences like turnout in
this precinct fell nine points, or this precinct swung twelve
to the challenger. They read like measurements, and each carries a
second claim that is never stated and almost never checked: that “this
precinct” is the same piece of ground in both elections. Can you
compare a precinct with itself four years later?
Often you cannot, and Houston County, Georgia (Warner Robins, Perry,
and 163,633 people in 16 precincts) shows why. Between the 2020 and 2024
presidential elections, 3 of its precinct names disappeared and a
precinct that had never existed appeared. Nothing in either year’s file
records what any change was: the 2024 map does not say what it replaced,
and no county has to announce that it moved a line.
The stakes are larger than one county’s bookkeeping. The precinct is
the smallest unit at which American voting can be observed, and
everything fine-grained is built on it: voting-rights analysis, campaign
targeting, journalism about which neighborhoods shifted. And counties
split precincts where population grows, so the places most likely to be
miscompared are exactly the fast-changing suburbs coverage cares about
most. An error of this kind does not look like an error. It looks like a
finding.
The test
Three files answer it, from three different publishers. The precinct
boundaries belong to Georgia’s Legislative and Congressional
Reapportionment Office, which posts a statewide precinct map for each
general election (https://www.legis.ga.gov/joint-office/reapportionment)
as a shapefile — the standard computer-readable map format, in which
each precinct is a shape with a row of data attached. The two read here are
the 2,684-precinct map of 2020 and the map that replaced it in 2024. The
population belongs to the Census Bureau: 232,717 Georgia census blocks —
the smallest areas the Bureau reports, often a single city block — from
TIGER/Line 2020, the Bureau’s own map files, carrying the 2020 census
count.
The votes belong to the Secretary of State’s certified precinct returns
for November 2020.
Why this data can answer: the precinct sits at the bottom of the
ladder of election reporting, the finest rung published statewide, and
the
levels-of-aggregation chapter walks that ladder. This brief asks the
question that chapter leaves open — whether the bottom rung holds still
between elections. Georgia is the test state because one office
publishes both years’ maps in one format, across the interval coverage
actually compares.
The data
A shapefile is not a picture: beside the geometry sits an attribute
table, one row per shape. The 2020 table:
| 1 |
ID |
row number inside the shapefile |
| 2 |
AREA |
the polygon’s area, in the shapefile’s own units |
| 3 |
DATA |
unlabelled — the name says nothing and the values do
not either |
| 4 |
DISTRICT |
a district number, which district is not stated |
| 5 |
CTYSOSID |
the Secretary of State’s county identifier |
| 6 |
PRECINCT_I |
precinct identifier |
| 7 |
PRECINCT_N |
precinct name — the other half of the manufactured
key |
| 8 |
CNTY |
county code |
| 9 |
FIPS2 |
a second FIPS-like code, differently padded |
| 10 |
CTYNAME |
county name — half of the key this chapter has to
manufacture |
| 11 |
REG20 |
registered voters, 2020 |
| 12 |
VOTED20 |
ballots cast, 2020 |
| 13 |
VOTED201 |
a second turnout column, distinguished only by the
digit DBF appended |
| 14 |
TRUMP20 |
votes for Trump, 2020 |
| 15 |
BIDEN20 |
votes for Biden, 2020 |
| 16 |
JORGENSON2 |
votes for Jorgensen — the name truncated at ten
characters |
| 17 |
PURDUE20 |
votes for Purdue, 2020 Senate |
| 18 |
OSSOFF20 |
votes for Ossoff, 2020 Senate |
| 19 |
HAZEL20 |
votes for Hazel, 2020 Senate |
| 20 |
LOEFFLER20 |
votes for Loeffler, 2020 Senate special |
115 columns, and the names are not the names anybody chose. The
format cuts every column name off at ten characters, so
JORGENSON2 is a truncated candidate. VOTED20
and VOTED201 are two different quantities, told apart by a
digit the format appended. The votes sit bolted straight to the
geometry; hold onto that. The 2024 release from the same office is a
different kind of object:
| 1 |
ID |
row number inside the shapefile |
| 2 |
AREA |
the polygon’s area, in the shapefile’s own units |
| 3 |
DATA |
unlabelled — the name says nothing and the values do
not either |
| 4 |
DATA1 |
a second unlabelled column |
| 5 |
DISTRICT |
a district number, which district is not stated |
| 6 |
CTYSOSID |
the Secretary of State’s county identifier |
| 7 |
FIPS |
county FIPS code |
| 8 |
FIPS2 |
a second FIPS-like code, differently padded |
| 9 |
CTYNAME |
county name — half of the key this chapter has to
manufacture |
| 10 |
CONTY |
sits immediately before COUNTY; appears to be a typo
that shipped |
| 11 |
COUNTY |
county |
| 12 |
PRECINCT_I |
precinct identifier |
| 13 |
PRECINCT_N |
precinct name — the other half of the manufactured
key |
13 columns and not one of them is a vote. Neither file carries a
field pointing at the other — no prior name, no successor code. The only
shared handle is county name plus precinct name, so the key — the label
on which rows of two tables are matched — has to be manufactured, and it
exists in no source file: HOUSTON|ROZR. About thirty rows a
year carry neither name and collapse to a single non-precinct,
NA|NA, which is dropped everywhere before anything is
counted.
One more caution sits upstream. The counties hold the official lines
and may redraw them at any time; the office itself calls its statewide
file a “best fit” of the 159 county maps, snapped to census
geography.
The census blocks are what make comparison possible. Each block
carries GEOID20, a fifteen-digit identifier the Bureau
maintains and does not reuse — a real key. Each block is reduced to a
point guaranteed to lie inside it, and that point is asked two
questions: which 2020 precinct contains you, and which 2024 precinct?
The answers are the block assignment: every block, with
the precinct it falls in under each year’s map. Three of ROZR’s
blocks:
| 131530212052000 |
169 |
HOUSTON|ROZR |
HOUSTON|ROZR |
| 131530212052001 |
26 |
HOUSTON|ROZR |
HOUSTON|ROZR |
| 131530212031010 |
12 |
HOUSTON|ROZR |
HOUSTON|PEC |
Two blocks stayed and one moved; that third row is this brief’s whole
finding, written out once. Grouping the 232,717 blocks by their pair of
precincts and summing population turns the assignments into a
crosswalk, a table that translates one set of areas
into another, with a weight on every pair.
crosswalk_pop.csv holds from_2020,
to_2024, the weight, and the pop
behind it; block_assign.csv holds every block’s
GEOID20, pop, and two precinct keys. The
running example:
| HOUSTON|ROZR |
HOUSTON|ROZR |
0.571018 |
11908 |
| HOUSTON|ROZR |
HOUSTON|PEC |
0.428982 |
8946 |
The votes are read straight off the 2020 map’s own attribute table
rather than joined on by name. That shortcut is legitimate only if the
map’s copy is right, so it was checked against the Secretary of State’s
certified returns:
| Georgia |
2,355 |
2,355 |
0 |
| Houston County |
16 |
16 |
0 |
On every precinct where the two publications could be compared, they
agreed to the vote. Matching the votes on by precinct name instead, as a
first attempt here did, silently lost 489,876 votes (9.8% of Georgia) to
spelling differences. The data was never wrong. The labels were.
Before you read on, commit to a number. Between 2020
and 2024, Houston County’s 16 precincts produced 4 changes: 3 names
disappeared and 4 name appeared that had not existed before. How
many of those changes moved actual ground from one precinct into
another, and how many are a polling place renamed while every
resident stayed put? Write down a number.
What it says about democracy
Start with the names, which are all most analyses ever see:
| Precinct names, 2020 |
16 |
2,651 |
| Precinct names, 2024 |
16 |
2,539 |
| Names present in both |
13 |
2,240 |
| Names gone by 2024 |
3 |
411 |
| Names new in 2024 |
3 |
299 |
| Share of names surviving (%) |
81.2 |
84.5 |
Only 84.5% of Georgia’s precinct names survive the four years, and
Houston, at 81.2%, is slightly harder than average. Measured on the
counties’ precinct codes instead of names, statewide survival
drops to about three quarters. Two keys, two answers, and neither is
wrong. An identifier that changes when the thing it names does not, and
holds still when the thing it names changes, is not an identifier.
Houston is small enough to show every change in full: every 2020
precinct and what became of it, then the new 2024 name, with the
precinct that supplied its people:
| ANNX |
ANNX |
unchanged |
| BMS |
BMS |
unchanged |
| CENT |
CENT |
unchanged |
| CGTC |
CGTC |
unchanged |
| FMMS |
FMMS |
unchanged |
| HAFS |
HAFS |
unchanged |
| HCTC |
VFW |
same ground, new name |
| HEFS |
HEFS |
unchanged |
| HHPC |
HHPC |
unchanged |
| MCMS |
MCMS |
unchanged |
| NSES |
NORTH HOUSTON SPORT COMPLEX |
same ground, new name |
| RECR |
WELLSTON CENTER |
same ground, new name |
| ROZR |
ROZR |
kept its name, lost part of its ground |
| TMS |
TMS |
unchanged |
| TWPK |
TWPK |
unchanged |
| VHS |
VHS |
unchanged |
| NORTH HOUSTON SPORT COMPLEX |
NSES supplied 100.0% |
| PEC |
ROZR supplied 100.0% |
| VFW |
HCTC supplied 100.0% |
| WELLSTON CENTER |
RECR supplied 100.0% |
Most readers guess that most of the changes are real. The
ground moved in 1 of the 4. The other 3 are renames: HCTC
became VFW, NSES became NORTH HOUSTON SPORT COMPLEX, RECR became
WELLSTON CENTER. In each case, weighting by population (counting some
blocks for more than others, by the people living in them) sends 100% of
the old precinct’s people to the new name. The polling place moved; the
ground did not. The one real change is ROZR, split in two, its western
half becoming the new precinct PEC — and there, two defensible ways of
dividing the old precinct’s votes disagree about 1,270 ballots.
Figure 1. Houston County’s precincts in 2020 and 2024, with
the two kinds of change separated. A pair of maps, because the
claim is about ground and only a map can show ground holding still while
names change. Amber areas are the 3 renamed precincts whose boundaries
stayed put; blue is ROZR, the one precinct whose ground actually
changed; gray kept both name and shape. The 2024 panel counts 17
precincts where the table lists 16; the missing one is PEC, explained
below.
The tempting repair, keeping only the precincts whose names match in
both years, is exactly wrong. In Houston, name-matching throws away 3
precincts (HCTC, NSES, RECR, carrying 31,161 people), and all three are
renames where nothing happened. It keeps ROZR, the one place where
something did. Name-matching is a filter on clerical churn, not on
boundary change, and a rule whose sign can flip between a county and its
state is not a rule to build on.
The fix is the crosswalk built above: newest boundaries as the
target, old data broken down onto blocks, blocks added back up inside
the new lines. A whole precinct reduced to one interior point can never
record a split, which is why PEC is invisible to the precinct-level
table; 3,269 blocks, each reduced to a point, can.
Then the choice that moves votes: when a precinct’s ballots travel to
more than one successor, in what proportion — by land, or by people?
| By area |
Share of the precinct’s land in each target |
15 of 16 (93.8%) |
78.3 |
| By population |
Share of the precinct’s people in each target |
1 of 16 (6.2%) |
18.5 |
The two rules disagree even about how many precincts split at all.
Most of area’s extra “splits” are slivers: hairline overlaps where the
two files traced the same road a meter apart. Raise the floor to 1% of a
precinct and area’s Houston figure falls from 93.8% to 6.2%, exactly
matching population. Statewide, a residue of 3.9 points survives the
correction: boundaries clipping empty farmland, a split in land that no
vote can come from. Where the choice bites is a real split:
Figure 2. The one piece of Houston ground that changed
precinct, and the 10,035 ballots that have to be divided.
ROZR’s 561 census blocks, each dot sized by population and colored by
where its interior point landed in 2024. The western half is big and
thinly settled, so land gives PEC 55.6% of ROZR while people give it
42.9%. In ballots: 10,035 × 55.6% = 5,575, against 10,035 × 42.9% =
4,305. The gap is 1,270 ballots, moved from one precinct to another by a
methodological choice, with nothing miscounted.
Dots on the dashed diagonal are pairs where land and people give the
same answer. Hover a dot for the pair and the gap.
Figure 3. Where land and people give different answers,
across Georgia. Each point is one 2020-to-2024 precinct pair,
placed by its weight under each rule, because the question is agreement
and a scatter shows agreement as distance from the diagonal. Of the
state’s 3,400 pairs, 324 (9.5%) differ by more than ten points of
weight. The labeled extreme, ELI WHITNEY COMPLEX in CHATHAM County, is
sent 11.0% of its precinct’s votes by one rule and 100.0% by the other.
Houston’s 17 pairs are green; its worst case is the ROZR split.
With a rule chosen, the carry itself is mechanical. It ends with the
comparison the opening asked for: Houston’s 2020 presidential vote
expressed on Houston’s 2024 map, including PEC, a precinct that did not
exist when those votes were cast:
| BMS |
4,812 |
2,685 |
10,481 |
7,637 |
| MCMS |
4,203 |
2,713 |
9,858 |
7,043 |
| FMMS |
3,439 |
2,983 |
9,567 |
6,551 |
| HHPC |
2,821 |
3,109 |
10,530 |
6,046 |
| ROZR |
3,605 |
2,029 |
8,481 |
5,730 |
| CENT |
3,130 |
2,353 |
8,366 |
5,584 |
| WELLSTON CENTER |
1,087 |
3,878 |
10,539 |
5,054 |
| VHS |
3,164 |
1,578 |
6,735 |
4,820 |
| PEC |
2,709 |
1,525 |
6,372 |
4,305 |
| VFW |
1,909 |
2,022 |
7,184 |
4,021 |
| CGTC |
2,236 |
1,374 |
5,182 |
3,680 |
| TMS |
1,493 |
2,098 |
5,995 |
3,657 |
| ANNX |
2,180 |
1,221 |
4,791 |
3,478 |
| TWPK |
1,868 |
1,328 |
4,757 |
3,278 |
| NORTH HOUSTON SPORT COMPLEX |
1,120 |
676 |
2,596 |
1,825 |
| HAFS |
1,083 |
299 |
1,891 |
1,391 |
| HEFS |
675 |
361 |
1,423 |
1,042 |
A crosswalk that loses votes is worse than none, because it looks
fine. So the carry is checked statewide against the totals it started
from:
| 2020 precincts (shapefile) |
2684 |
| carried into the crosswalk |
2653 |
| 2024 precincts receiving votes |
2696 |
| votes before |
13,821,357 |
| votes after |
13,821,335 |
| votes lost |
22 |
| share lost (%) |
0.000 |
22 votes lost out of 13,821,357. The trip can be audited, and this
one arrives whole.
What this brief cannot tell you
Population is not voters. The weights count children
and non-citizens, and turnout varies inside a precinct in ways a
headcount cannot see. Registered voters by block would be the right
weight, and nobody publishes it: that is the binding limit here.
The blocks are older than the boundaries. Block
populations are from the 2020 census; the 2024 lines came later, so
growth in between is invisible to the weights — in Houston, precisely
the growth that caused the split.
No number above carries an error bar. A carried-over
total prints with the same authority as a counted one, and in the
Houston table 2 of the 17 rows are estimates while nothing in the output
says which.
And one county is one county. Houston shows the
mechanism cleanly because it is simple. The statewide figures above say
the rate travels; the anatomy of any other county has to be read the
same way, by hand.
What you should have learned
- Only 84.5% of Georgia’s precinct names survived from 2020 to 2024,
and a changed name is usually not a changed boundary: Houston’s 4
changes were 3 renames and 1 real split.
- Keeping only name-matched precincts filters on the wrong thing: in
Houston it deletes 3 precincts where nothing happened and keeps the one
where something did.
- The weighting choice is not a technicality: land and people disagree
by more than ten points of weight on 9.5% of Georgia’s precinct pairs,
and in ROZR the choice moves 1,270 ballots.
- A carry can be checked: pushed through the population crosswalk,
Georgia’s 2020 vote arrives on the 2024 map with 22 votes lost.
Extensions
None of these is answered above, and every one can be answered from
the tables the Sources section hands over.
fig_twosource.csv compares the shapefile’s precincts
against the certified list. Find the precincts in one and not the other.
What is a precinct that exists on a map but not in the returns?fig_houston_lab.csv flags each Houston precinct as
kept, renamed or split. Count each kind. What share of the county’s
precincts would a name-matcher have compared correctly?fig_rozr_meta.csv divides ROZR’s ballots two ways and
reports both. Compare ballots_area against
ballots_pop. Which assumption would you defend to an
editor, and why?fig_rozr_blocks.csv assigns each of ROZR’s blocks a
target precinct and a pop. Find the blocks with no people
in them. What do empty blocks do to a weighting by area?crosswalk_pop.csv covers every county, with a
weight per precinct pair. Pick a county you know and count
its precincts that split. Quieter or louder than Houston?
Two stretch ideas point past the spreadsheet. The reapportionment
office posts precinct maps back to 2012; download an older pair and
measure whether this churn rate is new or normal. And find one county’s
own notice of a precinct change, in a board of elections agenda, and
check it against what the two maps imply.
Sources
Data
The data itself
Every figure above is drawn from these tables. They are plain CSV,
they open in a spreadsheet, and they are yours to take further.
fig_houston.csv, fig_houston_lab.csv,
fig_rozr_blocks.csv,
fig_rozr_meta.csv,
fig_rozr_outline.csv,
fig_rozr_pts.csv,
fig_twosource.csv,
crosswalk_pop.csv,
block_assign.csv
Method
