Skip to content

Burlington 2009 — the real election, the real ballots, three winners

The most-cited real-world IRV failure, reproduced in this repo from the actual 8,980 ballots (PrefLib dataset 00005, from Burlington's published records). One ranked-ballot set, three different winners depending only on the count:

→ Interactive visualization: Equal Vote's Real RCV — Burlington 2009 (a scrollytelling walk through the final rounds) · browse every real US RCV case study — most worked fine, which is the honest backdrop for why Burlington and Alaska 2022 stand out.

Count Winner
Choose-One Plurality (= round 1 first choices) Kurt Wright (R) — 2,950
RCV-IRV (the rule actually in force) Bob Kiss (P) — won the final 4,313–4,059
Any Condorcet count (Ranked Robin here) Andy Montroll (D) — beats everyone head-to-head

Every teaching claim on this page is runnable: the yamls below are the real profile, and Larry Hastings' open-source starvote engine, extended in this repo (the "LH" tabulator), reproduces the official figures within ±1 (a 6-ballot / 0.07% exclusion of equal-rank overvotes not representable in ranked syntax — every margin at stake is hundreds of votes).

What happened under IRV

ROUND 1                          SEMIFINAL                    FINAL
Wright     2950                  Wright     3293              Kiss    4313  Elected
Kiss       2585                  Kiss       2981              Wright  4059
Montroll   2062                  Montroll   2553  Rejected
Smith      1306  Rejected
Simpson      35  Rejected
WriteIn      36  Rejected

Montroll — the one candidate a majority preferred to each rival — had the fewest of the top three first choices, so IRV cut him in the semifinal. That is the center squeeze, not as a whiteboard construction but as a certified municipal result. (Run it.)

The pairwise truth

Round-Robin — every pair, head-to-head (For – Against):
   Montroll  beats Kiss       4063 – 3476
   Montroll  beats Wright     4596 – 3663
   Montroll  beats Smith      4569 – 2996
   Montroll  beats Simpson    6261 –  591
   Montroll  beats WriteIn    6656 –  100

    #  Candidate  W–L–T  Copeland  Margin
    1  Montroll   5–0–0         5  +15319   <- the Condorcet winner
    2  Kiss       4–1–0         4  +10742   <- the IRV winner
    3  Wright     3–2–0         3   +8850   <- the plurality leader

A perfect 5–0 record (run it). Wright voters preferred Montroll to Kiss 3-to-1 among those expressing a preference; Kiss voters preferred Montroll to Wright. The middle of the electorate agreed on the Democrat — and the count in force never asked the question.

The raise: Kiss gains 750 votes and loses

Non-monotonicity, built from the same real profile (run it). Let 750 Wright-first voters raise the winner Kiss to first — all 281 who already ranked Kiss second (their top two swap) plus 469 of the 840 Wright-only bullet voters (now Kiss > Wright):

SEMIFINAL                        FINAL
Kiss       3731                  Montroll  4063  Elected
Montroll   2553                  Kiss      3945
Wright     2543  Rejected

Kiss's new support knocks Wright below Montroll in the semifinal; Montroll survives to the final and beats Kiss. Kiss gained 750 first-place votes and lost the election he had won. More support, worse result — the additional-support paradox on certified ballots. (The selection is engineered but not exotic: the 469 bullet-ballot moves erode Montroll's 587-vote head-to-head cushion by only 469, so his final-round win survives the raise. The mirror reading — roughly the same bloc staying home also elects Montroll, an outcome those Wright voters preferred to Kiss — is the no-show paradox; Alaska 2022 repeated both patterns, worked in favorite_betrayal_voting_301.md §4.)

The spoiler test: delete Wright, and Montroll wins

The strict Independence-of-Irrelevant-Alternatives test is a thing you can run: strike a candidate who did not win out of every ballot, change nothing else, and see whether the winner moves. Strike Wright (run it) and it moves — from Kiss to Montroll, 4,063–3,476. The 840 Wright-only bullet ballots have nothing left to say and exhaust; nobody is handed a preference they did not express.

Those two numbers are exactly the Montroll-over-Kiss row in the pairwise table above. With Wright gone, the instant runoff simply is that head-to-head — the counterfactual invents no election, it asks the question the count in force never did.

This is the check behind FairVote's comparison chart, claim-checked: the chart rates RCV High on spoiler resistance because it "satisfies … the Independence of Irrelevant Alternatives … criteri[on]", and Burlington 2009 is one of the two elections FairVote itself names as its Condorcet failures. (Alaska 2022 is the other, and its pair runs the same way.) Note what this case is not: Wright (R) and Montroll (D) are not allies, so this is a plain IIA failure rather than vote-splitting between similar candidates — Alaska is the one that is both.

Reading this fairly

  • Kiss won under the rules everyone campaigned under, and IRV's defenders correctly note no method is paradox-free (Arrow — see the social-choice shelf). The claim this page supports is narrower and stronger: center squeeze, non-monotonicity, and no-show are IRV-specific mechanics — on these very ballots, Ranked Robin (and any Condorcet count) elects Montroll, and the paradox files above stop being constructible.
  • The voters did nothing wrong. Every pathology here operates on sincere ballots. Wright voters ranking honestly got their last choice; had they insincerely ranked Montroll first, they'd have done better — a real-world favorite-betrayal failure in the system marketed as making honesty safe.
  • Aftermath: Burlington repealed IRV in 2010. (It later re-adopted ranked ballots for city-council races in 2021 — the debate continues, which is exactly why the ballots deserve to stay runnable.)
  • Scope honesty: these are ranked ballots, so only ranked-ballot counts run natively; a STAR count would require invented score levels, and this repo doesn't invent data. The score-ballot version of this electorate's lesson lives in center squeeze with designed ballots.

Run it yourself

python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/burlington_2009/cases/burlington_2009_irv.yaml
python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/burlington_2009/cases/burlington_2009_ranked_robin.yaml
python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/burlington_2009/cases/burlington_2009_raise_kiss_nonmono.yaml

Full mirrors: IRV · Ranked Robin · the non-monotonicity replay. LH-only (no BetterVoting election — 8,974 weighted ballots across 378 patterns; the value here is fidelity to the record, not castability).

Sources. Ballot data: PrefLib dataset 00005 (00005-00000002.toi), derived from Burlington's published 2009 records. Official result figures (round 1 Wright 2951 / Kiss 2585 / Montroll 2063; final Kiss 4313–4061; pairwise Montroll over Kiss 4067–3477) reproduce here within ±1 after the 6-overvote exclusion. Analyses: Wikipedia, 2009 Burlington mayoral election (neutral overview); the non-monotonicity/no-show literature on Burlington and Alaska 2022 (Graham-Squire & McCune, arXiv:2301.12075). Sibling constructions in this repo: monotonicity pairs, participation/no-show pair, center squeeze.

file: README.md