Amherst College honors thesis · May 2026
Congressional redistricting is argued about as politics. I ask what it does to a city's money, and find that cities cut across two or more congressional districts collect measurably less federal funding per resident than otherwise similar cities that stay whole.
My honors thesis,
Economic Effects of Splitting Cities When Redistricting: Effects of Gerrymandering Beyond Politics
is considered unique because it bridges advanced data science, AI supported geospatial coding, and long-term economic tracking.
First-of-Its-Kind Database:
My thesis utilizes Python, ArcGIS, and Claude Code to create a first of its kind geospatial
database of congressional representation and fiscal outcomes over 50 years for all municipalities
across the continental United States.
Beyond Politics:
While most redistricting research focuses heavily on partisan wins and losses,
my work shifts the focus entirely to the local economic and fiscal aftermath when cities are fragmented by gerrymandering.
Agentic AI Integration:
My methodology stands out from traditional undergrad econometrics by aggressively utilizing
AI systems and coding tools to compile and synthesize half a century of local decentralized geographic data.
A representative whose district contains a whole city captures the entire electoral return from fighting for that city's federal grants. Split the city in two and each representative captures only half the return while paying the full cost in time and political capital. Each representative therefore has less incentive to push for federal funding. The split city therefore ends up with less advocacy in Congress and less federal funding relative to the size of its population.
The opposite prediction is also plausible: a split city has two members of Congress rather than one, sitting on two sets of committees, inside two coalitions. Both advocating and voting in the city's interests can raise the city’s influence and federal funding. This paper tests which mechanism is dominant, and by how much.
Each dot is one of the municipalities in the panel that can be placed on a map. Teal means the city sat inside a single congressional district that decade; orange means the district lines split the city. Drag the slider through seven redistricting cycles, or press play. Scroll to zoom, drag to pan, and click a dot for that city's record.
Type a name to pull that municipality's whole record: how many congressional districts it sat in each decade, what share of it fell in its biggest district, and what its federal transfers and local taxes per resident looked like in each Census of Governments wave.
Median federal intergovernmental transfers per resident, in constant 2017 dollars, for the municipalities in each state that report them. This is the raw data with nothing held constant — it is what the country looks like before any of the paper's controls, and the states differ for a hundred reasons besides district lines.
Medians, not means: a handful of large split cities with very high transfers pull the average around, which is exactly the problem the paper's fixed effects are there to solve. Split cities are also bigger and more urban than whole ones, so this picture is suggestive and nothing more. The 2022 reversal is the pandemic money — see the caveats.
Everything above is description. The paper's actual estimate comes from a two-way fixed-effects event study: it compares each city to itself over time, using the fact that redistricting moves lines around cities for reasons that have nothing to do with any one city's finances. 2002 is the reference year, so every point is measured against where that city stood in 2002. Points below the line mean less federal money.
The binary version — split or not — behaves the way the theory predicts: split cities sit below whole ones, significantly so in 2017. The continuous version measures the share of a city that falls in its largest district, so a higher number means a more unified city, and there the 2017 coefficient is negative — implying more unified cities got less. Rabin attributes that to a small number of heavily split large cities that receive enormous federal transfers for reasons unrelated to splitting, and notes that among the small and medium cities that make up almost all of the sample the raw relationship runs the other way. It is a real tension in the results, and the headline number rests on the binary specification.
Big cities are split all the time — nearly half of those above 100,000 people are — and it does not appear to cost them much: their representative still depends on them for votes. The penalty shows up in the small and mid-sized places, where a representative who serves half a town also serves a dozen others.
Event study coefficients on the share of a city in its biggest district, by 2002 population, baseline 2002. Only the coefficients the thesis reports are plotted; the large-city series for federal transfers is described as uniformly small and insignificant without figures, so nothing is drawn for it.
A second finding, and the one the paper itself treats most cautiously. Cities whose representative shares a party with both the governor and the president do worse, not better — the pattern the distributive politics literature calls being taken for granted. Every federal-transfer coefficient below is negative in 2012 and 2017, and every one of them flips positive in 2022.
The estimate is identified off the cities that change: those made whole and those cut apart at each cycle. Everything else is a control. Hover a bar for the counts.
Table 3 of the thesis, measured between the Census of Governments waves either side of each redistricting.
Applying the paper's headline figure — per resident per year — to a city of whatever size you like. This is arithmetic on a single estimate, not a forecast, and the estimate carries all the caveats below.
The whole design rests on state legislatures redrawing lines for reasons unrelated to any one municipality's finances. If mapmakers systematically split cities that were already on a different fiscal path, the estimate picks that up instead. The pre-2002 coefficients are the test, and for federal transfers they pass: cities always whole and cities always split were statistically indistinguishable in 1987, 1992 and 1997.
For local tax revenue the pre-period coefficients are negative and significant throughout, shrinking toward zero rather than sitting at it. Rabin reads that as a closing level difference rather than a diverging trend, which is plausible, but it means the tax revenue null is weaker evidence than the federal transfers result.
Every specification flips sign in 2022. The American Rescue Plan sent $65bn straight to municipalities in 2021, and the Infrastructure Act followed — one-off payments that ran through different channels for cities above and below 50,000 residents and that correlate with exactly the things being measured. The paper keeps 2022 in the main specification and drops it from the robustness check; the results hold without it.
The federal transfers result is concentrated in 2017 — 2012 is small and insignificant. Rabin reads this as fiscal effects taking a decade to show up after the 2010 lines were drawn, which is reasonable for grant cycles, but it does mean the headline rests on a single five-year snapshot.
Aligned and misaligned districts differ by region, by urban and rural, and by a dozen things no control absorbs. The paper says so plainly: the alignment channel cannot be cleanly separated from underlying structure the way the split channel can.
Districts drawn after the 2020 Census have had one Census of Governments wave to show up in, and that wave is the pandemic one. Whatever the current maps are doing to city finances will not be visible until the 2027 data lands. The paper says as much in its conclusion, and it is the main reason the 2022 estimates should not be read as the story reversing.
Linking Census of Governments financial records to municipal boundaries succeeds for 70 to 88 percent of the panel depending on the wave, and split cities match at a slightly higher rate than whole ones because they are bigger. The attrition is concentrated among small whole towns, which cuts against the finding rather than toward it.
Every coefficient, standard error and p-value on this page is transcribed from the PDF — the event studies, the size and partisan heterogeneity, the transition and match-rate tables, and the headline dollar figures. Nothing in those charts was re-estimated here.
The maps, the city records, the split-rate line and the split-versus-whole medians are built from her own data files: the places panel for split status and districts, and the Census of Governments extracts for money. They are plain descriptive statistics with no controls, and they exist to show the raw material, not to test anything.
Her panel carries no coordinates, so each municipality is placed using the Census Gazetteer: of them match. Her financial extracts identify governments three different ways depending on the wave — a real Census place code in 2017 and 2022, and Census of Governments codes before that — so the money is joined by identifier where possible and by exact name within state otherwise, dropped when the name is not unique. Coverage therefore varies by wave:
A count of five districts means five or more: the panel stores at most five per city, so Los Angeles and Chicago both read as five.
Built for Jason Furman from Maya Rabin's public repository. The paper is hers; the visualization, the joins and any mistakes in them are not.