# Likely-Voter Models A poll of adults measures opinion. A poll of *voters* predicts elections. Getting from the first to the second requires deciding which respondents will actually vote — and that decision is made by a model, not by observation, because the election has not happened yet. **The likely-voter screen typically moves results by more than the reported Margin of Error, and unlike the margin of error it is almost never reported.** ## How the screen is built Pollsters use some combination of: - **Self-reported intention.** "How likely are you to vote?" Wildly over-optimistic — far more people say they will vote than do. - **Past-vote history.** Self-reported (unreliable, people misremember and over-claim) or matched to the voter file (accurate but requires the panel be matched to registration records). - **Interest and enthusiasm items.** Attention to the campaign, ability to name the polling place. - **Demographic propensity.** Age, education, and residential stability are strong turnout predictors. - **Voter-file modeling.** Commercial vendors score every registered voter for turnout propensity from behavioral data. These get combined either as a **cutoff** (only respondents above a threshold count) or as a **weight** (everyone counts, in proportion to modeled probability of voting). ## Why it is the biggest hidden lever The screen and the weighting scheme together do more to determine a poll's topline than sample size does. Two competent pollsters given the identical raw responses can produce results several points apart purely through modeling choices — which electorate composition they assume, which turnout scenario they consider baseline. This creates a structural problem that is not about dishonesty: **The screen encodes an assumption about the electorate, and the assumption is drawn from the past.** A likely-voter model trained on prior cycles assumes the next electorate resembles them. When a campaign genuinely mobilizes irregular voters — or when a group's turnout collapses — the model is wrong in a direction it cannot detect from inside the poll. This is a large part of why polling errors in recent American cycles have been *correlated across pollsters* rather than random: everyone was making a similar compositional assumption, so everyone missed in the same direction, and the errors did not average out in aggregation. ## The political consequences **Turnout assumptions become self-fulfilling narrative.** A model assuming low youth turnout produces polls showing a race where youth do not matter, which shapes coverage, resource allocation, and expectations. **"Unskewing" is the abuse case.** In 2012 a cottage industry emerged around re-weighting public polls to a party-ID distribution the re-weighter preferred, producing forecasts of a Romney win. The methodological point — that weighting assumptions matter enormously — was correct. The application was motivated reasoning: the re-weighters selected the assumption that produced the desired answer rather than the assumption best supported by evidence. The episode is a clean case study because **the critique and the fallacy were the same observation, differing only in whether the conclusion was fixed in advance.** **Screens can carry demographic bias.** Screens built on registration stability, past-vote history, and residential permanence systematically down-weight young, low-income, and highly mobile voters — populations that also face the heaviest [[Voter Suppression]] friction. The model treats a suppressed voter and an uninterested voter identically, because from the model's perspective they look the same. ## What to ask of a poll Whether the sample is adults, registered voters, or likely voters; what the screen was; what electorate composition was assumed; and whether the pollster reports results under multiple turnout scenarios. Outfits that publish the last of these are being substantially more honest than outfits reporting a single number with a ±3 attached. ## Related Polling · Margin of Error · [[Poll Aggregation]] · [[Herding in Polling]] · [[Reading Probabilistic Forecasts]] · Statistics · [[Mathematics and Power Hub]]