How the ZIP Code Districts are drawn

This map redraws all 435 U.S. House districts by joining neighboring ZIP code areas. No voter registration, election results, party data, incumbent addresses or demographic data other than total population is used at any step. The same code, run on the same Census files, produces the same maps every time.

The map offers nine modes, each a different way to produce a first version of the districts. Every mode then goes through the same two finishing stages, so the modes differ only in how they start. The modes are listed from most to least defensible as a neutral rule; Ranking the modes explains the order.

• The building blocks
• The rules
• The shared finish: balancing and smoothing
• The nine modes
• Ranking the modes
• Results
• How the maps were checked
• Limitations and open questions
• Sources

The building blocks

ZIP Code Tabulation Areas. ZIP codes are mail delivery routes, not areas, so the Census Bureau publishes ZIP Code Tabulation Areas (ZCTAs). Each ZCTA is a block-based approximation of the land a ZIP code serves. This project uses the 33,585 ZCTAs in the 50 states from the 2020 Census cartographic boundary file. Throughout this page, “ZIP” means a ZCTA.

Population. Each ZIP’s population comes from the 2020 Census Demographic and Housing Characteristics file, table P1 (total population). That’s the same census that set each state’s number of seats.

ZIPs that cross state lines. 137 ZIP areas, home to 556,203 people, straddle a state line. State lines come first: a district can’t cross one, so each of these ZIPs is split at the state line into one piece per state (277 pieces, for example 42223 becomes a Kentucky piece and a Tennessee piece). Each piece’s population is the Census count for that ZIP “(or part)” within that state, from the same 2020 table. The pieces then work like any other ZIP. Every other ZIP stays whole, so the algorithm works with 33,725 building blocks: 33,448 whole ZIPs and 277 pieces. Pieces with no residents still go to a district in their own state, so every bit of a state’s ZIP land is covered.

Counties. Each ZIP (or piece) is assigned to the county in its state that holds most of its land, using the 2020 county boundary file. That assignment drives the Counties first mode and the “counties split” statistic.

Who is left out. The District of Columbia and the territories have no voting House seat, so they are excluded.

Neighbors. Districts must be connected, so the algorithm needs to know which ZIPs are neighbors. Two ZIPs are neighbors when:

1. Their boundaries touch, allowing 25 meters for small gaps in the boundary file (88,190 pairs).
2. They face each other across empty land. ZIP areas don’t cover unpopulated land: national forests, deserts, parks, military ranges. ZIPs that border the same uncovered area (at least 1 km²) and lie within 25 km of each other are treated as neighbors (9,432 pairs). A district may legally span empty land, and without these links places like El Paso would be attached to the rest of Texas by only a thread.
3. Water is the only link. Islands are connected to the nearest ZIP in the same state (140 links).

Map colors. District colors carry no meaning; they only tell neighbors apart. Every map uses exactly five colors, each on about the same number of districts (87), and districts whose borders touch get different colors, including across state lines. Four colors are always enough in theory, and reference maps conventionally use a handful of soft colors so borders and labels stay readable. In a few modes, New York City districts interlock so tightly that seven or eight of them all touch one another. There, at most four pairs that touch only through one or two ZIP codes share a color, and the district border still divides them. The map offers eleven color themes: Accessible (the default), Standard, Factbook, Atlas, Terrain, Greens, Bold, Poster, Graphite, Lunar and Blueprint. Accessible uses Paul Tol’s “light” palette and Bold uses the Okabe–Ito palette, both designed to stay distinguishable with color-vision deficiency.

Distances are measured in equal-area map projections: one for the lower 48, and separate Albers projections for Alaska and Hawaii.

Why ZIP codes: prior research

Political scientists John Curiel and Tyler Steelman have studied what happens when district lines cut through ZIP codes. Current congressional districts split the ZIP codes of over 100 million Americans. In survey data from 2008 to 2016, people living in split ZIP codes were less likely to recognize or contact their representative, and felt more distant from them ideologically. Most people look up their representative by ZIP code, and several members of Congress ignore mail from outside their district. A follow-up study found that people in split ZIP codes were also less likely to know the race of their representative, which weakens the purpose of districts drawn under the Voting Rights Act. Curiel and Steelman proposed a zero-tolerance standard for ZIP splits. Their North Carolina simulations showed that whole-ZIP maps are feasible and substantially reduce partisan bias.

This project builds on that idea in a different way: a fixed, politics-blind rule applied to every state, with the code and outcomes published. In every mode here, each ZIP belongs to exactly one district, so a ZIP code tells you who your representative is. The only exception is the 137 ZIPs a state line runs through, where the ZIP and your state together tell you.

The rules

“The rules” means the complete recipe the program follows. Written down in full, it lets anyone who runs it on the same Census files get the same map. Every item below changes the result, so whoever chooses them holds real power. What keeps that power from becoming a gerrymander is when the choices are made (see why the rules are fixed in advance).

Constraints every map follows

• Each state keeps its 2020 seat count. Districts never cross state lines.
• Equal population within each state. The national average is 761,169 people per seat. Districts can’t cross state lines, though, so each district aims for its own state’s ideal: the state’s population divided by its seats. That’s about 542,000 in Montana and about 990,000 for Delaware’s single seat.
• Whole ZIPs, except at state lines. A district is a set of complete ZIP areas. State lines take priority: a ZIP that a state line runs through is split at that line, and nowhere else.
• Contiguity. Every district is one connected piece, using the neighbor rules above.
• No political inputs. Only geography, population and county lines.

Six states have one seat (Alaska, Delaware, North Dakota, South Dakota, Vermont, Wyoming). There the whole state is the district, and those states are left out of the deviation statistics.

The complete recipe

1. Building blocks. The ZIP Code Tabulation Areas published with the census, frozen at that moment (2020: 33,585 in the 50 states), with the 137 that cross a state line split at it. ZIP codes the Postal Service adds or retires later are ignored until the next census, so nobody can move a district by lobbying for a new ZIP code.
2. Population. Total population per ZIP from the same census (2020 DHC, table P1).
3. Seats and target. Each state’s apportioned seats. Target = the state’s population ÷ its seats. Deviation should be as small as whole ZIPs allow (currently about ±1–2%). Best score treats ±2% as its eligibility cutoff.
4. Neighbors. Boundaries touching within 25 m; ZIPs facing each other across at least 1 km² of uncovered land within 25 km; islands linked to the nearest ZIP in the state. See the building blocks.
5. Method. One mode, applied the same way in every state.
6. Shared finish. Balancing (up to 400 passes), then smoothing: 600 steps per ZIP, border weight μ = 0.000004, or in the county-aware modes μ = 0.00003 with cuts along county lines at 15% of the cost. Fixed random seed. See the shared finish.
7. Tie-breakers. Written into each mode, for example “the cut closest to north–south” in the splitline modes.
8. The code. The published source is the recipe in a form a computer can run. Given the same files, it reproduces every map exactly.

Why the rules are fixed in advance

Gerrymandering means looking at possible outcomes, then choosing the map you like. Fixing the rules means choosing the recipe before anyone can see what it produces. Is choosing the rules just another way to gerrymander? Only if the rules are picked after looking at the partisan results. These safeguards prevent that:

• Adopt the rule now, apply it to the next census. The 2030 data doesn’t exist yet, so nobody, in either party, can know whom the rule will help.
• Use one rule for all 50 states. A single nationwide rule can’t be tailored to help one party: it would help in some states and hurt in others.
• Choose on stated principles, before any partisan analysis. Publish the reasons for each choice, then measure the outcome.
• Pick one mode. This map shows nine modes for comparison. In a real redistricting cycle, offering a menu would hand the choice to whoever picks the mode for each state. That would just move the gerrymander from drawing lines to picking modes. Ranking the modes makes the case for one.
• Publish random seeds after the rules are locked. This only matters for the random modes; see Random draw.

Fair process, not engineered outcomes

This project judges fairness by the process, not by the result. Districts follow people: a dense city gets exactly as many districts as its population earns, and the map neither knows nor cares who lives in a ZIP.

There is a serious counter-view. When one party’s voters are packed into cities, that party wins those seats by large margins, and its extra votes there earn it nothing more. Its share of seats can then fall below its share of votes even on a politics-blind map. People who judge fairness by outcomes argue that maps should be drawn to correct for this, which requires voter data. This project doesn’t do that. It does hold that partisan outcomes should still be measured and published after the maps are drawn, so the trade-off is visible to everyone (see Limitations).

The shared finish: balancing and smoothing

Every mode produces a first draft. Two stages then improve it, identically for all modes.

Stage 2: balancing. The algorithm visits every ZIP on a district border. It moves the ZIP to the neighboring district if that lowers the total squared population deviation of the two districts and leaves both in one piece. It repeats until no move helps (at most 400 passes). A small tie-breaker prefers giving a ZIP to the district whose center is closer.

Stage 3: smoothing (simulated annealing). Balancing alone gets stuck. Imagine a district that is 3% over, but every ZIP it could give away is too large or would cut it in two. Annealing explores more freely. It scores the whole state as:

score = Σ (district deviation)² + μ × (cut edges)

A cut edge is a pair of neighboring ZIPs that end up in different districts. Fewer cut edges means smoother, more compact borders, and this is a standard compactness measure in redistricting research. μ is small (0.000004), so population equality dominates. Border smoothness acts as the tie-breaker.

Each step picks a random cut edge and proposes moving one of its ZIPs across. Improvements are always accepted. Worse moves are sometimes accepted early on and almost never at the end. That lets the search escape dead ends, then settle. If moving a ZIP would cut off a small piece of its district (under 15% of a district’s population), the piece moves with it, which lets districts pass through narrow rural corridors. The run lasts 600 steps per ZIP in the state, and the best scoring map seen is kept.

The random numbers in this stage come from a fixed seed, so reruns are identical.

The nine modes

The modes fall into three groups:

• Fixed rule modes apply one mechanical rule.
• Start point modes grow districts outward from an anchor.
• Pick by rule modes let a rule set before any map existed make the choice: one random draw from a public seed, or the best-scoring of the other modes’ maps.

They are described here in ranking order, the same order as on the map (why).

Shortest splitline

In one sentence: cut the state with the shortest possible straight line, then cut each half the same way.

Warren D. Smith proposed this rule specifically to take choice out of redistricting.

1. For a region that needs k districts, split it into two parts holding A = ⌊k/2⌋ and B = k − A districts. For example, 9 seats split 4 : 5.
2. Try every straight cut direction in 1° steps (180 directions). For each direction, sort the ZIPs by position across the cut and slide the cut until one side holds as close to A/k of the region’s population as possible. Each ZIP goes to the side its population center falls on.
3. Measure each candidate cut’s length, and keep the shortest. Ties go to the cut closest to north–south.
4. Repeat inside each half until every piece needs one district.
5. Then the shared finish, which repairs any piece disconnected by a cut and fine-tunes population.

Adapting splitline to ZIPs. The original rule measures a cut across the state’s outline. Here, a cut’s length is measured across the populated ZIPs nearest the cut (the closest 3% of the region’s ZIPs, at least 8). Empty land between ZIP areas doesn’t count toward length, so cuts favor narrow populated crossings. Because whole ZIPs are kept, district borders follow ZIP boundaries near the cut instead of exactly straight lines.

Strengths: no settings at all, so only one possible map. Compact, simple shapes. Weakness: straight cuts ignore communities, rivers and county lines.

Splitline + counties

In one sentence: shortest splitline, but the pieces being cut are whole counties.

1. As in Shortest splitline, a region that needs k districts is split A : B (A = ⌊k/2⌋), trying every cut direction in 1° steps.
2. For each direction, counties are lined up by their population center along that direction (the ZIPs inside a county by their own position). The cut slides along this line-up until one side holds A/k of the population. So every county lands whole on one side, except the single county where that share is reached, which is divided along the straight line.
3. The resulting border follows county lines, so its length is measured as the number of neighboring ZIP pairs it separates (cut edges). The shortest border wins. Ties go to the direction closest to north–south.
4. Repeat inside each piece. Inside a county big enough for several districts (Cook County, Los Angeles County), every ZIP is in the same county, so the cut is a plain splitline.
5. Then the shared finish, with the county-aware settings borrowed from Counties first: cutting along a county line costs 15% of cutting through a county, μ = 0.00003, and a state beyond ±2% gets one more pass at the normal weight.

Result: the fewest split counties of any mode (615 nationwide) and the second smoothest borders (16,030 cut edges, behind only Best score). Iowa shows the difference: 7 split counties, the same as Counties first, versus 26 for Shortest splitline. Cost: the least even population of the fixed rules: average deviation 0.38%, largest 2.71% (NY-5 in Queens, where ZIPs are dense and large), 393 of 429 districts within ±1%. Strength: keeps splitline’s single rule while respecting the places people know. Weakness: one more rule (counties are the units) and the two county settings in the finish.

Counties first

In one sentence: keep counties whole wherever population allows, as Iowa’s nonpartisan agency does.

1. Growth works like NW corner, with one change: a district finishes the counties it has already entered before it may enter a new county.
2. In the smoothing stage, cutting a border along a county line costs about 15% of what cutting through a county costs, and borders carry more weight overall (μ = 0.00003). So borders settle on county lines where they can.
3. If any district in a state ends up more than ±2% from ideal, that state gets one more smoothing pass at the normal border weight. This happens in New York, where each NYC borough is a county with several districts’ worth of people.

Result: about 40% fewer split counties than the start point modes (670 nationwide versus roughly 1,100–1,120), at the cost of slightly less even population. Strength: districts line up with places people already know. Weakness: county sizes vary enormously, so perfect county integrity is impossible in large states.

State center

Works like NW corner, except the anchor is the middle of the state’s extent (the center of its bounding box, not its population center). District 1 forms around the middle and later districts form around it. Districts at the edges are drawn last, so they absorb whatever is left over and tend to be less tidy.

NW corner

In one sentence: fill each state like reading a page, starting from its northwest corner.

1. The anchor is the northwest corner of the state’s extent.
2. District 1 starts at the populated ZIP closest to the anchor.
3. The district grows one neighboring ZIP at a time. It always picks the candidate with the lowest value of distance to the district’s population-weighted center + 0.35 × distance to the anchor. The first term keeps districts compact. The second sweeps the filling away from the corner, so no stray gaps are left behind.
4. Growth stops at the state’s ideal population. The last ZIP is added only if that leaves the district closer to ideal than stopping short.
5. Any pocket of unassigned ZIPs too small to hold a district (under half the ideal) is absorbed into a neighboring district.
6. The next district starts at the unassigned ZIP closest to the anchor. The last district takes whatever remains.
7. Then the shared finish.

Strengths: simple, explainable, deterministic. Weakness: the choice of corner is arbitrary, and the NE corner gives a visibly different map.

NE corner

Identical to NW corner, except the anchor is the northeast corner of the state’s extent. Comparing the two shows how much the starting corner matters.

ZIP number

In one sentence: let the Postal Service’s numbering decide the order.

1. District 1 starts at the state’s lowest-numbered populated ZIP.
2. It grows by always adding the neighboring ZIP with the lowest number, until it reaches the ideal population.
3. The next district starts at the lowest-numbered unassigned ZIP.
4. Then the shared finish.

ZIP numbers aren’t random. The first three digits identify a mail-sorting center, and numbers within a region roughly follow delivery geography. So districts in this mode tend to follow postal regions. Strength: uses an existing, politically neutral numbering. Weakness: postal regions were designed for mail, not representation, and the resulting shapes can be irregular.

Best score

In one sentence: for each state, take the best map the other eight modes produced, judged by a rule fixed in advance.

1. Candidates: the eight other modes’ maps for that state.
2. Eligibility: keep maps whose largest district deviation is within ±2% (if none qualify, all stay eligible).
3. Winner: the fewest cut edges, meaning the smoothest, most compact borders.
4. Tie-breakers: lowest average deviation, then the ranking order. In one-seat states every mode gives the same map, so the top-ranked mode, Shortest splitline, “wins” there.

The map shows which mode won each state. Nationally the winners spread across all eight other modes: Splitline + counties wins 19 states, Shortest splitline 15 (6 of them one-seat states), Random draw 6, Counties first 5, NE corner 2, and State center, NW corner and ZIP number 1 each. This mode has the fewest cut edges overall (14,464) and the third fewest split counties (735).

Why fixing the rule first matters: if the scoring rule were chosen after looking at the maps, choosing the rule would just be a way of choosing the map.

Random draw

In one sentence: one random map per state, from a seed nobody could choose.

Randomness is only fair if nobody can redraw until they like the result. This mode allows exactly one draw.

1. Seed. Each state’s seed is a hash (FNV-1a) of the text 2020 Census apportionment population 331108434|<STATE>|random. The number is the official 2020 apportionment total, published before any of these maps existed. Anyone can recompute the seed and get the identical map.
2. Random spanning tree. Give every neighbor link in the region a random weight and take the minimum spanning tree: a set of links that connects every ZIP with no loops.
3. Split. Removing any one link of a tree splits it into exactly two connected pieces. Find every link whose removal leaves one side within ±3% of A/k of the population, and pick one at random. If no tree has such a link, draw new trees (up to 400) and keep the closest.
4. Recurse into each piece until every piece needs one district.
5. Then the shared finish.

This “random spanning tree” method is the standard way researchers sample possible districtings, because both sides of every cut are guaranteed to be connected.

Caveat: the apportionment total was published before these maps, but it was also known when the rules were written. For a real redistricting cycle, the seed should be a public value released after the rules are locked in, such as a lottery draw or a public randomness beacon.

Ranking the modes

The question: how many choices does each mode leave that a person could tune? Every such choice is a lever, and a lever can be pulled toward one party. Fewer choices means less room for bias. The balance and border statistics are reported in Results, but they don’t decide the rank: a mode that needs tuning to look good isn’t neutral.

Recommendation. Shortest splitline is the most defensible as a neutral rule. Splitline + counties is the strongest choice if keeping counties whole matters most. It adds one rule and two settings, but those settings are plain and fixed, and it splits the fewest counties of any mode. Random draw isn’t less fair than the modes above it; it ranks low because it is harder to explain and to trust.

Results

Deviation is measured against each state’s own ideal, over the 429 districts in multi-seat states.

The 3.36% cases are the same district in New York (one district made of the same 17 ZIPs), and Splitline + counties’ 2.71% is also in New York (NY-5, Queens): large, dense ZIPs and island geography limit how finely population can be traded there.

Partisan outcomes (measured after drawing)

The maps are drawn without any political data. After they were finished, each mode was scored against the 2024 presidential vote, as promised under Fair process, not engineered outcomes. These numbers were not used to choose, tune or rank any mode, and the ranking above is unchanged.

How it was measured. Precinct results from the New York Times 2024 precinct map were assigned to ZIP areas by area overlap, then added up for each district. Where the precinct file is incomplete (mainly Oklahoma and Tulsa counties in Oklahoma), official county totals fill the gap, spread by ZIP population, so Oklahoma’s district results are approximate. Presidential votes stand in for each district’s partisan lean. They are not House results, and real House races also turn on candidates, incumbency and turnout.

Nationally, Harris won 49.15% of the two-party vote. A seat share matching that vote share would be about 214 of 435 seats.

Efficiency gap: the difference in each party’s “wasted” votes (votes for a loser, or a winner’s votes beyond 50%), as a share of all votes. Median district vs. nation: how much redder the middle district is than the country. 50/50 vote: every district shifted by the same amount until the national vote is tied. The enacted-map row is The Downballot’s count for the districts used in 2024.

What the numbers show. Every mode gives Harris fewer seats than her vote share would suggest, by about 10 to 22 seats, and every mode’s median district is redder than the nation. The enacted 2024 maps show a similar tilt. Most of the gap comes from a few states: under shortest splitline, Harris falls furthest short of a proportional share in Florida, Texas, Ohio, Georgia, North Carolina, Indiana and Pennsylvania, and does better than proportional in California, New Jersey, Massachusetts, Connecticut and Washington. The modes differ by up to 12 seats from each other. Splitline + counties and Best score lean furthest toward Republicans, NE corner and Counties first the least. ZIP number produces the most competitive seats (129), with Splitline + counties, Best score and Shortest splitline close behind.

Responsiveness: how seats follow votes

A map is responsive when a shift in how people vote produces a shift in who wins. To measure it, every district’s 2024 presidential result is moved by the same amount until the national two-party vote reaches a given level, and the seats are counted again.

Under shortest splitline, each 2-point shift in the national vote moves about 21 to 30 seats. At any given vote share, the nine modes differ by 8 to 14 seats. The national vote moves far more seats than the choice of mode. Near a 50–50 vote, though, the mode can decide which party holds the majority.

Voters represented by someone they voted against

In any system of single-member districts, everyone who voted for a losing candidate is represented by someone they voted against. This number shows how much the map itself changes that.

*For actual House races: Democratic and Republican House voters whose representative is from another party.

About 4 in 10 voters end up represented by someone they voted against, whichever map is used. The ZIP maps raise that share by only 1–1.5 points over the enacted maps, about 2 million voters. That is the cost of having more competitive seats: when more races are close, more voters end up on the losing side of one. The map does change who is on the losing side. In every case, Democratic voters lose more often than Republican voters, because more of them live in lopsided city districts.

The actual House figure is lower partly because 40 races had no Democrat-versus-Republican contest, so many voters there had no real choice to lose.

How to read it. People who judge maps by process will see this as political geography at work: Democratic voters are concentrated in cities, so blind, compact districts waste more of their votes, and no one chose that result. People who judge maps by outcome will see a consistent tilt that a fair map should correct. This analysis doesn’t settle that debate. It makes the trade-off visible, and both readings start from the same numbers. One election is also a narrow test. The result depends on where each party’s voters live, which shifts over time.

The per-district, per-state and per-mode results are in data/outcomes/, and the scripts are pipeline/06a–06c.

How the maps were checked

A separate program, written independently of the districting code, checks every mode and confirms that:

• every one of the 33,725 building blocks (33,448 whole ZIPs and 277 pieces of ZIPs split by a state line) is assigned to exactly one district;
• every block’s district is in that block’s own state;
• no ZIP is divided except by a state line: the only divided ZIPs are the 137 a state line runs through, one piece per state;
• each state has exactly its number of seats, for 435 districts in total;
• every district is one connected piece;
• district populations add up to the total population.

All nine modes pass.

Limitations and open questions

• Exact population equality. Courts hold congressional districts to near-exact equality (Karcher v. Daggett, 1983). Whole ZIPs can hold up to about 130,000 people, so ±1–2% is near the practical limit for a ZIP-only map. Exact equality would mean splitting a few ZIPs along census blocks.
• Voting Rights Act. A mechanical method doesn’t check compliance with Section 2, which can require districts that give minority communities an opportunity to elect their preferred candidates.
• ZIP areas aren’t ZIP codes. ZCTAs approximate delivery areas, and some addresses fall in a ZCTA with a different number than their mailing ZIP.
• Blind to politics is not the same as neutral in outcome. Researchers Jowei Chen and Jonathan Rodden argue that because one party’s voters are more concentrated in cities, compact, politics-blind maps can systematically favor the other party. Others, including Michael McDonald, dispute how strong that effect is. This project judges fairness by process: no political data, rules fixed in advance (see Fair process, not engineered outcomes). Partisan outcomes are measured after drawing and published in Partisan outcomes: every mode tilts toward Republicans on 2024 presidential results, by about as much as the enacted maps.
• One census. Everything is based on 2020 data. The pipeline can be rerun after the 2030 census.
• Seed for the random modes. See the caveat under Random draw.

Sources

• U.S. Census Bureau, 2020 Census Demographic and Housing Characteristics File, Table P1 (total population), by ZCTA.
• U.S. Census Bureau, 2020 cartographic boundary files: ZCTAs (cb_2020_us_zcta520_500k), counties (cb_2020_us_county_500k), states (cb_2020_us_state_500k).
• U.S. Census Bureau, 2020 Census apportionment results.
• Warren D. Smith, shortest splitline algorithm (rangevoting.org).
• Iowa Legislative Services Agency redistricting criteria (Iowa Code §42.4).
• DeFord, Duchin and Solomon, “Recombination: A family of Markov chains for redistricting” (random spanning-tree methods) (mggg.org).
• Chen and Rodden, “Unintentional Gerrymandering: Political Geography and Electoral Bias in Legislatures,” Quarterly Journal of Political Science, 2013.
• The New York Times, “An Extremely Detailed Map of the 2024 Election,” precinct results (non-commercial use with credit) (github.com/nytimes/presidential-precinct-map-2024).
• Tony McGovern, 2024 county-level presidential results (github.com/tonmcg/US_County_Level_Election_Results_08-24).
• The Downballot, 2024 presidential results for all 435 House districts (the-downballot.com).
• MIT Election Data and Science Lab, U.S. House 1976–2024 returns (doi.org/10.7910/DVN/IG0UN2), and 2024 presidential results by congressional district compiled by Drew Savicki (via jaytimm/PresElectionResults).
• Curiel and Steelman, “Redistricting Out Representation: Democratic Harms in Splitting Zip Codes,” Election Law Journal 17(4), 2018, 328–353 (preprint); summarized in “Here’s one way to end partisan gerrymandering: Don’t break up Zip codes,” The Monkey Cage, 2018 (themonkeycage.org).
• Steelman and Curiel, “Redistricting Out Descriptive Representation: The Harmful Effect of Splitting ZIP Codes on the Constituent–Representative Link,” Journal of Race, Ethnicity, and Politics 8(1), 2023 (cambridge.org).