American Multi-Party Electoral Simulation
A full-stack simulation of what American politics might look like under proportional representation, using real ideological data from the 2024 Cooperative Election Study (CES): 60,000+ respondents, 45,707 after listwise deletion.
▶ Explore the live simulation: https://usmultiparty.com/
Methodology: docs/METHODOLOGY.md — how 45,707 survey responses become a ten-party legislature, with caveats.
For methodology, see docs/METHODOLOGY.md; for EFA factor loadings, docs/EFA_FACTORS.md; for which file is authoritative for a given number, docs/DATA_SOURCES.md. Agents start at CLAUDE.md.
Overview
This project explores what American politics might look like under different electoral rules:
- Multi-party proportional representation instead of winner-take-all
- Single Transferable Vote (STV) primaries with Droop quota surplus transfers
- A rolling presidential primary with geographic balance across 4 rounds
- Ranked Pairs Condorcet and IRV for senate general elections (reported side by side)
- Two candidate fields — party-line vs. crossover — run through every chamber and the presidency (the crossover field is dormant: still in the pipeline, hidden from the site)
- Ballot depth — how many candidates voters actually rank (3–10 or all; default 7): short ballots exhaust, fill STV seats below quota, and center-squeeze the consensus winner out of the finalist field
Crossover field (dormant)
The crossover field is hidden from the site. It is the least data-driven half of the model — the variants
are synthetic positions (one axis shifted by ±25% of the inter-party SD) scored by factor-space proximity
rather than by the DPGMM posterior everything else uses — and it never picked up the bootstrap, the turnout
stops, or the ballot-depth variants, so it falls further behind the party-line side with each change.
Nothing is deleted: the pipeline scripts, the factor_deviation/ outputs, the JSON payloads, and every
crossover-only card in the viz still exist and still work. viz/src/constants/features.ts only removes the
affordances that reach them. Two ways back in:
- load any page with
?lab=crossover — read once at startup, so it survives tab navigation and lasts until reload
- flip
CROSSOVER_BY_DEFAULT to true in that file and rebuild, to restore it for everyone
Regenerating the field is unchanged — see the crossover block under Running the Pipeline.
District map
data/processed/county_to_district.csv assigns each of the 3,142 counties to a multi-member
district, and it is an input to the House simulation — STV pools voters district by district.
draw_geographic_districts.py used to draw a different map on every run, because tie-breaks
resolved through set iteration order, which varies with PYTHONHASHSEED. That is fixed: the map is
byte-identical across runs, verified at seeds 0, 1 and 7.
The committed map predated the fix, so it could not be reproduced from the code. Resynced on
2026-07-31: the deterministic map is now the committed one, and everything downstream was rebuilt
against it. 65 counties changed district — all within their own state and their own density tier,
so no state's seat count moved — reshaping 31 of the 149 districts and reassigning three seats
(Conservative +1, Nationalist +2, Labor −1, Populist −2). Six parties and the 873-seat total were
unaffected.
pipeline/resync_district_map.py remains the driver for doing this again: six stages, dry-run by
default, --run to execute. Verified during the resync — senate, president and primary blocks in
uncertainty*.json came back byte-identical, which is the expected result since none of those
contests are district-pooled, and it confirms the bootstrap is reproducible given its seeds.
Measured cost: ~16 min per turnout stop, ~110 min for the seven, ~2 h end to end.
No-Solidarity scenario (removed 2026-09-11)
A robustness check that dissolved Solidarity (cluster 2) and let its voters' ballots flow to the
remaining parties. Its UI toggle was added, removed, restored and removed again, and nothing in the
app ever imported its payloads.
It is gone: the NO_STY=1 branches in pipeline/pure_only/, build_nosty_scenario(), the
pure_multi_nosty* output trees and the viz/src/data/archive/*NoSTY*.json payloads were all
deleted. To resurrect it, restore from history rather than re-adding an environment branch.
Pipeline at a Glance
2024 CES Survey (N=45,707)
│
▼
Exploratory Factor Analysis (EFA)
24 policy items → 5 latent ideological dimensions (F1–F5)
│
▼
DPGMM Clustering
Factor scores → 10 voter typology clusters (C0–C9)
│
▼
Two candidate fields (used for House, Senate, and President)
• Party-line (pure_multi): 3 identical-platform candidates per party
• Crossover (factor_deviation): candidates shifted ±1 axis off base
│
├──────────────────────────────────────┐
▼ ▼
House STV Simulation Senate Simulation
873 seats / 180 districts 51 senators (one per state)
Droop quota + Gregory surplus STV primary → Condorcet + IRV
→ pure_multi/ + factor_deviation/ 4 scenarios (2 fields × 2 methods)
│ │
└──────────────┬───────────────────────┘
▼
Chamber Vote Model
38 bills (stv_config.BILL_VARS)
Sum-of-Binomials → P(pass)
The 10-Party Typology
Produced by Dirichlet Process Gaussian Mixture Model (DPGMM) clustering on 5 EFA factor scores. All 10 clusters are active parties, including OAO (cluster 7, the Order & Opportunity Party — the former "Blue Dogs," a law-and-order Democratic bloc kept as its own party). DISSOLVED_PARTIES = [].
| ID | Abbrev | Name | Character | ~Electorate |
|---|
| C0 | CON | Conservative | Mainstream center-right; pro-law enforcement, economically traditional | ~19% |
| C1 | SD | Social Democrat | Center-left institutionalist; supports safety net, moderate on social issues | ~16% |
| C2 | STY | Solidarity | Disaffected working-class left; skeptical of institutions, economically left but low union density | ~15% |
| C3 | NAT | Nationalist | Populist right; strong on immigration, anti-establishment, very high F5 | ~9% |
| C4 | LIB | Liberal | College-educated progressive; socially liberal, moderate-left economics | ~9% |
| C5 | POP | Populist | Right-of-center reformists; skeptical of elections, high F2 + F5 | ~11% |
| C6 | CUP | Civic Union Party | True centrists; cross-pressured, low electoral skepticism | ~10% |
| C7 | OAO | Order & Opportunity | Law-and-order Democrats (former "Blue Dogs"); a full party that wins seats | — |
| C8 | DSA | DSA | Progressive left; far left on security, high electoral skepticism | ~6% |
| C9 | PRG | Progressive | Progressive elite; urban, far left across most dimensions | ~5% |
House seat counts (party-line, canonical, 873 total): CON=202, LBR=158, STY=129, POP=104, LIB=93, CUP=89, NAT=43, DSA=25, OAO=15, PRG=15. Conservative is the largest party. Source: viz/src/data/houseSeats.json — see docs/DATA_SOURCES.md. (Do not quote clusterProfiles.json seatsHouse for seat counts — it's a population baseline, not seats won.)
The 5 Ideological Factors (EFA)
Factor scores are standardized to the survey population (mean≈0, SD≈1). Absolute tier thresholds reflect position relative to the full electorate:
| Tier | Score Range |
|---|
| Very High | > +0.75 |
| High | +0.25 to +0.75 |
| Medium | −0.25 to +0.25 |
| Low | −0.75 to −0.25 |
| Very Low | < −0.75 |
F1 — Security & Order
High = pro-law enforcement, pro-border security, pro-surveillance
Low = civil libertarian, anti-enforcement
Top items: support increased police (+0.73), increase border patrols (+0.71), deny asylum (+0.66), oppose decreasing police (+0.65)
F2 — Institutional Distrust
High = distrusts elections and government; anti-establishment
Low = trusts institutions
Top items: state/local elections not fair (+0.90), US elections not fair (+0.73)
Anchored by the election-integrity items, but it also absorbs the government-trust items (CC24_423/424) and carries the real distrust signal — F2 correlates +0.84 with raw government distrust across the ten parties. Near-orthogonal to partisan ID — STY, POP, and DSA all score High despite being ideologically distant on F1/F5, so institutional distrust cuts across the left-right divide. (Formerly labeled "Electoral Skepticism.")
F3 — Government Distrust (residual)
Top items: distrust federal govt (+0.66), distrust state govt (+0.48)
Non-interpretable residual; the label is misleading and the sign inverts. Re-deriving from the stored factor scores (repro: analysis/efa/exploration/verify_f3_inversion.py), a party's F3 score correlates −0.38 with its actual government distrust: the two most distrusting parties, POP and NAT, score lowest, while institutionalist CON scores highest among the right. The mechanism is that the government-trust items (CC24_423/424) load on both F2 and F3, and F2 — where the party-level variance sits — absorbs the real signal, so F3's discriminating variance instead comes from culturally-coded cross-loadings (immigration, race) and its ordering tracks a culturally-progressive residual, not distrust. F3 earns its k=5 slot only by keeping the trust variance split so the DPGMM can isolate CUP, the institutionalist that trusts both elections and government (see Robustness in docs/EFA_FACTORS.md). It is excluded from party-facing displays and flagged as a residual; the genuine distrust axis is F2. All 23 winning coalition types also score Medium on F3 (range −0.21 to +0.13), so it does not differentiate coalitions.
F4 — Religious Traditionalism
High = traditional religious values; conservative on abortion and same-sex marriage
Low = secular, socially progressive
Top items: church attendance (+0.69), abortion week limits (+0.69), oppose same-sex marriage recognition (+0.65)
F5 — Populist Conservatism
High = populist-right; anti-immigration, fiscal conservatism, racial traditionalism
Low = progressive-left
Top items (negative-loaded): racial resentment (−0.62), oppose police reform (−0.56), oppose Dreamers (−0.54), oppose tax hike on $400k+ (−0.53)
NAT is the extreme high end (+1.51); PRG (−0.99) and LIB (−0.95) are the extreme low end.
Caveats
Institutional distrust reads as a floor, not a level. Clustering requires a complete answer on all 24 items, which restricts the analysis sample to the ~45,700 respondents who returned for the post-election wave. That attrition is not random: the returners run about eleven years older and markedly more politically engaged than the full 60,000, and the CES post-election weights correct the demographic margins but not the engagement gap. The exposure concentrates on one battery. The institutional-trust items, whether elections are run fairly and how far federal and state government can be trusted, are all post-wave, so the voters who left are absent from exactly the questions their attitudes would move most. And among those who did answer, the least engaged are the most distrustful: they score highest on election-illegitimacy and on distrust of state government. The alienated, low-information, "system is rigged" voter is the one least likely to have stayed to say so. So the simulation more likely understates institutional distrust than overstates it, especially toward elections and state government.
Broader modeling caveats (party cohesion, sincere voting, static factor space, turnout) live in docs/METHODOLOGY.md and the simulation's Caveats tab.
House STV Simulation
Scripts: pipeline/stv_main.py and supporting pipeline/stv_step1.py–pipeline/stv_step5.py
Published result: data/outputs/pure_multi/house/stv_seat_summary.csv → viz/src/data/houseSeats.json (the party-line view the viz shows).
Note: published seat counts come from pure_multi. The outdated 850-seat canonical/stv_seat_summary.csv was deleted 2026-09-12, along with three other CSVs in those directories that nothing read. The canonical* directories are retained for their inputs rather than their results: the pure_multi and factor_deviation runs read ballots_checkpoint.parquet + district_apportionment.csv, and the viz transfer matrix is built from canonical/transfer_matrix_10party.csv. See docs/DATA_SOURCES.md.
- 873 seats across 180 multi-member districts (Urban / Suburban / Rural tiers per state)
- Apportionment: Hamilton method, ~380,000 pop/seat from 2020 Census
- Droop quota:
⌊ total_weight / (seats + 1) ⌋ + 1
- Gregory surplus transfer: fractional weight redistribution when a candidate exceeds quota
- Scenarios may dissolve a party (
pre_dissolved); the baseline dissolves none. A dissolved party's ballots skip to the next-ranked active party.
- Ballots derived from DPGMM soft cluster probabilities via Plackett-Luce ranking
Seat results (party-line, canonical):
| Party | Seats | % | Urban | Suburban | Rural |
|---|
| CON | 202 | 23.1% | 107 | 67 | 28 |
| LBR | 158 | 18.1% | 89 | 51 | 18 |
| STY | 129 | 14.8% | 72 | 49 | 8 |
| POP | 104 | 11.9% | 52 | 37 | 15 |
| LIB | 93 | 10.7% | 63 | 24 | 6 |
| CUP | 89 | 10.2% | 47 | 31 | 11 |
| NAT | 43 | 4.9% | 16 | 16 | 11 |
| DSA | 25 | 2.9% | 15 | 8 | 2 |
| OAO | 15 | 1.7% | 9 | 4 | 2 |
| PRG | 15 | 1.7% | 11 | 3 | 1 |
This is the party-line field, the only one the site shows. The crossover field (run_fd_house_stv.py → data/outputs/factor_deviation/house/) runs the same STV over axis-shifted candidates; the House tab's Scenario toggle switches between the two fields, and appears only when the crossover flag is on.
Senate Simulation
Scripts: crossover field — pipeline/run_fd_senate_simulation.py; party-line field — pipeline/pure_only/run_pure_multi_senate.py (each runs both Condorcet and IRV)
Outputs: data/outputs/factor_deviation/senate/ and data/outputs/pure_multi/senate/ — senate_composition.csv (Condorcet) and senate_irv_composition.csv (IRV)
- 51 senators — one per state (50 states + DC).
- The senate uses the same two candidate fields as the rest of the simulation (there are no "blend" candidates):
- Party-line (
pure_multi) — 3 candidates per party, all on the identical party platform (e.g. STY_1). Tests proportional voting with parties as monoliths.
- Crossover (
factor_deviation) — 9 pure party candidates + 28 crossover candidates that shift one party off its base on a single axis: Security & Order, Anti-Establishment, or Populist Conservatism (±). Variant codes like STY_lo_ae (Solidarity, less anti-establishment).
- In each state, an STV primary narrows the field; the general is then decided two ways — Ranked Pairs Condorcet (beats-all winner) and IRV (last-place elimination).
- Two fields × two methods = 4 scenarios.
Senate seats by party across the four scenarios:
| Party | Crossover Cond | Crossover IRV | Party-line Cond | Party-line IRV |
|---|
| STY | 34 | 19 | 33 | 11 |
| SD | 11 | 27 | 15 | 26 |
| CON | 1 | 3 | 1 | 11 |
| POP | 5 | 1 | 1 | 2 |
| CUP | 0 | 1 | 1 | 0 |
| LIB | 0 | 0 | 0 | 1 |
NAT, DSA, and PRG win 0 senate seats in every scenario. Pattern: Condorcet rewards broad acceptability, so Solidarity (the spatial median) dominates; IRV rewards first-choice intensity, so Social Democrat leads and Conservative breaks through in the party-line field. Full state-by-state results are in comms/party_profiles.md.
Chamber Vote Model
Script: pipeline/chamber_vote_model.py
Outputs: data/outputs/senate/senate_vote_model.csv, data/outputs/house_vote_model.csv
For each of 37 binary CC24_ policy items, models the probability of a bill passing a floor vote.
Method — Sum-of-Independent-Binomials → Normal approximation:
E[Y] = Σᵢ nᵢ · pᵢ (expected yes votes)
σ²[Y] = Σᵢ nᵢ · pᵢ · (1 − pᵢ)
P(pass) = 1 − Φ((majority − 0.5 − E[Y]) / σ[Y]) (continuity correction)
where nᵢ = seats held by type i, pᵢ = that type's % supporting / 100.
Verdict thresholds: PASS ≥67% | TOSS-UP 33–67% | FAIL ≤33%
Results (both chambers): 29 PASS / 1 TOSS-UP / 7 FAIL
These chambers are cross-cutting — they simultaneously pass both tax cuts AND tax hikes, both border enforcement AND Dreamer protections. Different majority coalitions form for each vote. This is expected behavior for proportional multi-party representation.
Retired: blended senator types
An earlier senate model elected blended senator types — CON/CUP, STY/LBR, SD/LIB and the
like — each a weighted mix of two clusters. The current senate elects pure party types only, so the
whole layer is gone from the live pipeline and from the app. Six scripts moved to
pipeline/archive/: run_senate_simulation.py and run_senate_irv.py (the mixed senate itself),
generate_blend_stats.py and senate_chamber_profile.py (its profiles), and
cross_chamber_coalitions.py and analyze_senate_ideology.py (the coalition-type diagnostics).
chamber_vote_model.py lost its senate half for the same reason.
Their outputs are deleted too: data/outputs/senate/, data/outputs/coalitions/ and
data/outputs/profiles/blend_stats.csv. generate_results.py was the only remaining reader — it
charted the mixed senate to an untracked results/ directory nothing links to — so it moved to the
archive with them.
Output Files Reference
All outputs are under data/outputs/.
| File | Description |
|---|
pure_multi/house/stv_seat_summary.csv | Party-line House seat totals by party + density tier (→ houseSeats.json) |
factor_deviation/house/stv_seat_summary.csv | Crossover House seat totals (→ fdHouseSeats.json) |
pure_multi/senate/senate_composition.csv · senate_irv_composition.csv | Party-line Senate — Condorcet / IRV winner per state |
factor_deviation/senate/senate_composition.csv · senate_irv_composition.csv | Crossover Senate — Condorcet / IRV winner per state |
canonical/{ballots_checkpoint.parquet, district_apportionment.csv} | Shared ballot + apportionment cache the pure_multi / factor_deviation runs read as input — see DATA_SOURCES.md |
canonical/transfer_matrix_10party.csv | Vote-transfer matrix (→ viz transferMatrix.json) |
profiles/cluster_stats.csv | Per-item statistics for all 10 clusters (policy + demographics) |
senate/senate_chamber_profile.csv · house_chamber_profile.csv | Seat-weighted policy aggregates per chamber (vote-model inputs) |
senate/senate_vote_model.csv · house_vote_model.csv | 37-item bill passage probability per chamber |
coalitions/coalition_type_profiles.csv (+ _factor_alignment, _pairwise) | Secondary cross-chamber factor-alignment diagnostic |
Running the Pipeline
All scripts live in pipeline/ and use relative paths anchored to the project root.
All 10 clusters are active parties (C7 = Order and Opportunity Party). The election
sims + prepare_data.py are the reproducible path; they read data/processed/*.csv and
the geography checkpoint in data/outputs/canonical/ (party-independent).
# ── Base geography checkpoint (already cached; rebuilds ballots + apportionment) ──
python3 pipeline/stv_main.py
# ── Party-line field (pure_multi): ballots, primary, president, house, senate ──
python3 pipeline/pure_only/generate_pure_multi_ballots.py
python3 pipeline/pure_only/generate_party_ballots.py # single-candidate ballots (transfers)
python3 pipeline/pure_only/run_pure_multi_primary.py
python3 pipeline/pure_only/run_pure_multi_presidential.py
python3 pipeline/pure_only/run_pure_multi_house_stv.py # → pure_multi/house/
python3 pipeline/pure_only/run_pure_multi_house_stv.py --triple # → pure_multi_triple/house/
python3 pipeline/pure_only/run_pure_multi_senate.py # → pure_multi/senate/ (Condorcet + IRV)
# ── Crossover field (factor_deviation) ──────────────────────────────────────
python3 pipeline/generate_factor_deviation_candidates.py # OAO fields base only (small party)
python3 pipeline/generate_factor_deviation_ballots.py
python3 pipeline/generate_factor_deviation_stats.py # per-candidate policy support
python3 pipeline/run_fd_house_stv.py # → factor_deviation/house/
python3 pipeline/run_fd_house_stv.py --triple
python3 pipeline/run_fd_senate_simulation.py # → factor_deviation/senate/ (Condorcet + IRV)
python3 pipeline/run_fd_primary_2028.py
python3 pipeline/run_fd_irv_2028.py # FD presidential
# ── Canonical house + shared profiles ───────────────────────────────────────
python3 pipeline/run_house_canonical.py # → canonical/ (party-card seats)
python3 pipeline/house_chamber_profile.py # House chamber policy aggregate
# ── Respondent-level extras (read the raw .dta directly) ────────────────────
python3 pipeline/add_compare_items.py # FP/abortion/voting items → cluster_stats.csv
python3 pipeline/compute_intensity.py # full multi-point distributions → viz/src/data/clusterIntensity.json
# ── Visualization ──────────────────────────────────────────────────────────
cd viz && python3 scripts/prepare_data.py # Regenerate all JSON from outputs
cd .. && python3 pipeline/build_viz_bundles.py # REQUIRED after prepare_data — see below
cd viz && npm run dev # Dev server
npm run build # Production build → dist/
build_viz_bundles.py is not optional. Nine derived artifacts sit downstream of
prepare_data.py: the depth bundles the presidency, primary and house tabs lazy-load, the
rank-7 senate and legislation families that overwrite what prepare_data.py just wrote, the
single-race voter file, the top-two instructive-ballot splits, and the current-party profiles.
Skip it and the app renders numbers from the previous run, or 404s a bundle and silently drops
a whole section — nothing warns you. Run --list to see the order, --dry-run to see the
commands. Every step is idempotent, so rerunning the whole thing is always safe.
The two hex-tiling builders (build_hex_ec_cartogram.py, build_hex_seat_cartogram.py) are
deliberately not in it: the tilings key off apportionment rather than the per-run turnout and
depth data, and the EC one needs one invocation per basis (--basis electoral,
--basis population). Rebuild those by hand when seat counts change.
Legacy senate vote-model chain (senate_chamber_profile.py → chamber_vote_model.py
→ generate_results.py, plus cross_chamber_coalitions.py) produces analysis CSVs and
depends on the archived single-candidate pure_only/senate/ run. It is not required
for the viz: prepare_data.py recomputes every Legislation column (Party-Line, Crossover,
Triple — House and Senate) directly from the native seat counts × per-cluster policy
support, so those numbers reflect the current 10-party chambers without the old chain.
Data provenance / regeneration notes
EFA/clustering use the full ordinal responses — the 0/1 collapse is display-only.
The typology is built from a polychoric correlation matrix (analysis/efa/pipeline/run_efa.py)
and Bartlett factor scores from the standardized full-scale item matrix
(analysis/efa/pipeline/efa_update.py): multi-point items enter at full resolution
(reverse-coding only flips direction; "Not sure" on the 1–3 trust items is mapped to the
scale midpoint, not dropped), so neutral/middle opinions are preserved in the factor scores
and cluster assignments. Binarization to a single "% agree"/"% increase" happens only in the
display layer (add_compare_items.py); compute_intensity.py recovers the full distributions
for the viz. (An earlier commit message on the Intensity view wrongly stated the clustering
ran on collapsed encodings — it does not.)
stv_main.py and generate_blend_stats.py were archived in an earlier cleanup and have
been restored to pipeline/ and re-keyed to the current party codes. The only piece still
archive-only is the single-candidate pure_only/senate/ run, which feeds the legacy
senate vote-model chain (see the note above) and is superseded by prepare_data.py's
recompute. Older archived copies under data/outputs/archive/ and pipeline/archive/
predate three party renames (Reform→Populist, Center→Civic Union, SD→Labor, C7→OAO) and
should not be restored as-is.
Dependencies
python3 >= 3.10
numpy
pandas
scipy
scikit-learn
pyreadstat # Stata .dta reading
pyarrow # Parquet I/O for ballot checkpoint
plotly # HTML cluster profile visualizations
openpyxl # Excel read for county density-tier map
Data Sources
| Dataset | Description | Path |
|---|
| 2024 CES | ~60,000 respondents, 2024 Cooperative Election Study | data/raw/2024 CES Base/CCES24_Common_OUTPUT_vv_topost_final.dta |
| Typology | DPGMM cluster assignments (45,707 rows) | data/processed/typology_cluster_assignments.csv |
| EFA scores | 5 factor scores per respondent | data/processed/efa_factor_scores.csv |
| Polychoric matrix | 24×24 polychoric correlation matrix | data/processed/polychoric_matrix.csv |
Analysis sample: N=45,707 after listwise deletion (24 items + commonpostweight non-missing).
