Against the lean

How the ground is measured

A state’s lean is its presidential margin relative to the nation, averaged over the two most recent presidential elections before the race — the same construction as a Cook PVI, built here from the presidential returns already in this project. A race breaks the lean when the state leaned at least 5 points one way and the other party won it. The universe is every US Senate race from 2002 onwards plus every governor’s race in our poll panel: 667 races, of which 106 broke the lean and 68 did so on ground leaning at least 10 points the other way. (See How Close Is Close for what a margin like that means in practice, and why this page's line sits where it does relative to the site's other thresholds.)

A lean is not a prediction. It says how a place votes for president, which is the only baseline available this far back for every state; a Senate or governor’s race can differ from it for reasons that have nothing to do with either candidate.

Most of it is not a swing at all

Of the 68 seats won on hostile ground, 45 were the same party keeping a seat it already held and only 19 actually changed hands (4 could not be resolved). That distinction matters, because a party holding a seat in hostile territory is usually one person’s standing with their neighbours rather than anything the opponent did. The eight most extreme are all of that kind: long-serving incumbents and governors who run far ahead of their party’s presidential vote, year after year.

“Distance travelled” is the result minus the lean. It is deliberately crude: it mixes in the national mood of that year, so treat it as a rough size, not a measurement of a person.

The seats that really did swing

These 19 changed hands against a lean of 10 points or more. Read the last two columns together: in almost all of them the polls had the eventual winner ahead already.

When the competitor really was the problem

6 of those flips are set out below, each one a case where a source we fetched attaches a specific difficulty to the favourite — a conviction, an indictment, a scandal, a recall question, or an unpopular record. Every claim is a quote under 300 characters with a link. This is not a census of the 19 flips: it covers the cases we sourced, and a seat missing from it is a seat we did not research, not a seat where the favourite was fine. Phase one measured that bias directly, and it is large.

Look at what the polls did. In 4 of the 6 the poll average already had the winner ahead — including Alaska in 2008, where Ted Stevens was convicted eight days before the vote and the polls had still moved to Mark Begich. A damaged favourite is damaged in public, and the polling is taken afterwards. The two the polls still got wrong are Alabama in 2017, where the allegations broke about four weeks out and the average moved but not far enough, and Maryland in 2014, where there was no scandal at all.

The polls were mostly right. The map was wrong.

Across all 65 of these races that were polled, the poll average named the eventual winner in 52 of them. Of the 13 it got wrong, 10 had a poll average inside three points anyway — the polling said the race was a coin flip and the coin came down against the lean, which is not the same as a polling failure. In only 3 was the eventual loser clearly ahead.

So “a candidate swung the state” is almost always a statement about the map being a poor guide to a particular race, not about the polls missing something. That is worth saying plainly, because the two get confused constantly.

What this does to our own model

Hypercane, this project’s model for the risk that polls have named the wrong winner, works by shrinking a poll margin toward the state’s lean. On ground that votes against its lean, that is the one thing the model cannot afford to trust — and it shows.

Where states vote with their lean the model is close to a solved problem: it catches 84% of the races where the polls named the wrong winner while crying wolf on only 6% of the rest. Where states vote against their lean it catches 22% and raises false alarms on 32%. It gets worse at the job and noisier at the same time. The gap in catches is 61.6 points (95% confidence interval 30 to 90, Fisher exact p = 0.0011, on 31 and 9 poll misses respectively).

This is a blind spot with a mechanism, not bad luck. It also explains a loose end from part one, where Hypercane missed the upsets of Doug Jones, Larry Hogan and Susan Collins: all three happened on ground that votes against its lean. The honest reading is that the lean is load-bearing for this model and it is the wrong thing to lean on in exactly the races people most want called.

Sample sizes here are small — 9 poll misses on lean-breaking ground against 31 elsewhere — and the confidence interval is wide. The direction is solid; the precise size is not. Nothing has been changed in the model on the strength of this, and the operating point stays where it is until it is signed off.

What would break these numbers

Four things, stated because they are the ones most likely to matter.

Governors are only here if they were polled. Senate races come from the complete returns series; governors come from the poll panel, so the count of hostile races leans on which races attracted polling. The poll comparisons are unaffected, since they can only use polled races anyway.

A two-party margin misreads a strong three-way race. Maine in 2010 shows a 27-point result because the winner took the seat with well under half the vote against an independent; Vermont in 2002 is similar. The returns series carries a third-party share and the poll panel does not, so governors cannot be screened for this.

Races an independent won are excluded outright. Joe Lieberman’s 2006 re-election is filed in the source data as a 61-point Democratic win, and Angus King’s two Maine victories read as Republican ones. Left in, Maine 2012 would appear here as a state voting 40 points against its own lean.

Whose seat it was is inferred. A Senate seat’s previous holder is taken from the same state six years earlier and a governor’s from the previous election there, which leaves 4 of the hostile races unresolved and would misread a seat whose earlier race was uncontested.

Sources

Results and presidential baselines from the MEDSL returns already in this project; poll averages from its own panel; the 6 cases from the reports listed on each card, as fetched on 2026-09-16. The Alabama 2017 special election is added by hand with its own citation, because the returns series covers even years only. Parts one and two of this research are Long Odds and Safe Seats.