Glossary

Plain-English definitions

40 polling and political terms used on this site.

A B C E F G H I K L M O P R S T W

Approval rating: The percentage of respondents who say they approve of an officeholder's job performance. Reported alongside disapproval; net approval = approve − disapprove.
Ballot measure: A proposition appearing on a ballot for direct voter decision (initiative, referendum, constitutional amendment, bond, or legislatively-referred question).
Bellwether: A county, state, or demographic that historically votes with the eventual winner. Reliability fades in polarized cycles.
Brier score: The mean squared error of a probabilistic forecast: average of (predicted probability − outcome)², where outcome is 1 if the event happened, 0 if not. 0 is perfect, 0.25 is a coin flip; lower is better.
Class of senator: Each US senator belongs to one of three rotating classes (1, 2, 3). Only one class is up for election in any given cycle. 2026 is Class II.
Crosstabs: Subgroup breakdowns of a poll (by age, race, education, region). Often noisy because subsample sizes are small.
ECE (expected calibration error): A scalar summary of calibration: the average gap between predicted probability and observed frequency across probability bins, weighted by how many forecasts fall in each bin. 0 means perfectly calibrated; lower is better.
Electoral College: 538 electors total who formally elect the president — 535 apportioned to the states (each state's allocation = senators (2) + house seats) plus 3 for the District of Columbia under the 23rd Amendment.
Ensemble forecast: PoliAgg's headline model — a stacked ensemble that blends several base models (a Kalman polling filter, a gradient-boosted model, ridge, random forest, Cook PVI, market-implied odds, rater consensus, and fundraising) into one win probability and median margin per race. It's the number the map colors and the race-page hero show.
Favorability: Whether respondents view a person favorably or unfavorably, separate from voting intent.
Generic ballot: A poll question asking which party respondents will support for Congress, without naming candidates. Best read as a directional national signal.
House effect: A pollster's persistent tendency to overstate one side. Computed by comparing each pollster's polls to the broader consensus.
Incumbency advantage: The historical electoral edge sitting officeholders enjoy from name recognition, fundraising, and constituent service. Has narrowed in recent cycles.
Independent expenditure (IE): Money spent to support or oppose a candidate by a group not coordinating with the campaign (e.g. a super PAC). Reported to the FEC on Schedule E and shown in the money-in-race section.
Kalman filter: A state-space model that estimates a latent “true” value from noisy observations. Used here as one of the six aggregation methods.
Likely voter (LV) screen: A pollster's method for filtering registered voters down to those most likely to actually vote. Methods vary widely.
LOESS: Locally weighted scatterplot smoothing — a non-parametric regression that produces a smooth trend line less sensitive to outliers.
Margin of error: The 95% confidence interval around a poll's reported number, scaling with √(1/n). For a typical 600-person poll: ±4 pt.
Mean absolute error (MAE): The average size of a pollster's miss — the mean of |poll margin − actual result| across the races we have outcomes for. Lower is more accurate; it's the basis for the pollster accuracy grade.
Monte Carlo: A simulation method that draws thousands of random samples from each race's probability distribution to estimate chamber-level outcomes.
Out-of-fold (OOF): A prediction made for a race the model never saw while training — produced by holding that race's cycle out, fitting on the rest, then scoring the held-out cycle. Prevents the model from grading its own homework.
Oversample: Deliberately surveying more respondents from a small subgroup to improve subgroup precision; weighted back to its true share for the topline.
Partisan pollster: A pollster affiliated with a campaign, party, or political organization. Their polls are stored but flagged separately and weighted carefully.
Pollster accuracy grade: A letter grade (A+ to F) summarizing a firm's empirical accuracy, derived from its mean absolute error on the final poll it ran in each completed race. Firms with no completed-race track record yet show NR (not rated).
Posterior: In Bayesian estimation, the updated belief about a quantity after combining a prior expectation with observed data. Our recalibrated rating probabilities are posteriors: the rating-table value (prior) shrunk toward what history actually showed.
Primary vs general: Primary election polls measure intra-party preference; general election polls measure D-vs-R matchups. Methodologies differ.
PVI: Cook Partisan Voting Index — how much a district leans left or right of the national average across the last two presidential cycles.
Rating shift: When a race rating changes — e.g. Tossup → Tilt D. Tracked for both poliagg and external raters.
Registered voter (RV) screen: A poll of registered voters, before applying any likely-voter screen. Tends to slightly favor Democrats vs LV samples.
Right track / wrong track: A standard question asking whether the country is on the right or wrong track. Predictive of incumbent-party performance.
Safe seat: A race where the rating is Safe D or Safe R — incumbent or party advantage exceeds 15 points.
Split-ticket: When a voter chooses candidates from different parties for different offices (e.g. R for President, D for Senate). Has declined since 2010.
Tipping-point race: The race that delivers the decisive seat (or electoral vote) when contests are ordered by margin — the one most likely to determine which side wins control.
Topline: The headline number from a poll — the candidate-vs-candidate result before crosstabs. What gets quoted in the news.
Tossup race: A rating used when polls and fundamentals indicate a margin under 0.5 points. Closest possible bucket on the 9-tier scale.
Tracking poll: A poll repeated frequently (often daily) by the same pollster, with rolling samples to detect short-term shifts.
Turnout: The share of eligible voters who actually cast ballots. The biggest cycle-to-cycle source of margin variation.
Two-way margin: The difference between the top two candidates' shares (D% − R%), ignoring third parties and undecideds.
Weighted average: A polling average where each poll is weighted by sample size, recency, and pollster accuracy — vs an unweighted mean.
Win probability: The model's estimated chance a candidate or party wins, on a 0–100% scale. Distinct from margin: a 1-point lead in a stable race can carry a higher win probability than a 3-point lead in a volatile one.