How Presidential Polls Work

Every poll begins with a sample — a subset of people intended to represent a larger population. For a presidential poll, the target population is typically all likely voters in the United States, roughly 130-150 million people in a presidential election year. A pollster cannot interview all of them, so they interview a manageable number — typically between 600 and 1,500 respondents — and use statistical theory to generalize the results.

The key principle is random sampling: every person in the target population should, in theory, have an equal chance of being selected. In practice, this is extremely difficult. The era of random-digit telephone dialing — when nearly everyone had a listed landline — made true random sampling feasible. Today, with declining landline ownership, caller ID avoidance, low response rates (often under 5%) and a shift to online panel polling, the "random" in random sampling is increasingly theoretical.

Most modern polls draw from one of several sources: registered voter lists (which are public records), commercial consumer databases, opt-in online panels (where people have signed up to take surveys), or text message outreach. Each method has biases. Online panels oversample the internet-engaged. Registered voter lists miss newly registered voters. Phone polls underrepresent people who do not answer unknown numbers. Managing and correcting for these biases is the core challenge of modern polling.

Registering to vote and actually voting are two different things. In presidential elections, roughly 60-65% of eligible adults vote — which means any sample drawn from registered voters or all adults includes millions of people who will not cast a ballot. A poll that includes non-voters in its results will produce meaningfully different numbers than one that filters them out.

Likely voter (LV) models are the set of criteria pollsters use to determine which respondents to count in their reported results. The two most common approaches are:

• Self-reported likelihood: Simply ask respondents "How likely are you to vote in the upcoming election?" and include only those who say they are "very likely" or "certain to vote." This is simple but susceptible to social desirability bias — people who say they will vote but do not.
• Composite scoring: Score respondents on multiple factors — past voting history (derived from public voter files), enthusiasm, interest, registration duration and stated intention — and include only those above a threshold score. Gallup pioneered this method with a seven-question scale. It is more predictive but more expensive.

The choice of likely voter model can shift a poll result by 2-4 percentage points. A pollster who expects high Republican turnout will build a more Republican-leaning likely voter screen; one expecting Democratic enthusiasm will tilt Democratic. This is one reason different pollsters show different results even when interviewing at the same time.

Even with a well-designed sample, the people who respond to a poll are never perfectly representative of the electorate. Certain groups — older people, more educated people, politically engaged people — are more likely to respond to polls than others. If a poll's raw results include 65% college-educated respondents but college graduates make up only 38% of the electorate, the results are biased.

Weighting corrects for this by mathematically adjusting how much each respondent's answers count. A respondent from an underrepresented group gets a weight greater than 1 (counts more); a respondent from an overrepresented group gets a weight less than 1 (counts less). Pollsters typically weight simultaneously on age, gender, race, education, geographic region and party registration.

The critical — and controversial — decision is what the "correct" target looks like. For a registered voter poll, the target can be the actual registered voter population. But for a likely voter poll, the pollster must estimate what the electorate will look like on Election Day — a future event with inherent uncertainty. If a pollster assumes the electorate will look like 2020 but it ends up looking like 2016 (with higher white working-class turnout), their weighting targets will be wrong and their results will be systematically biased. This is the proximate cause of most major polling errors.

The margin of error (MOE) is the most commonly cited and most commonly misunderstood statistic in polling. It is calculated from sample size using standard statistical formulas: a poll of 800 respondents produces a margin of error of roughly ±3.5%; a poll of 1,600 respondents produces a margin of roughly ±2.5%. Quadrupling the sample size only halves the margin of error — which is why large-sample polls are expensive but not proportionally more accurate.

At a 95% confidence level, a stated margin of error means: if you ran the identical poll 100 times with different random samples, 95 of those polls would produce a result within the stated range. A candidate polling at 47% with a ±3% MOE is actually anywhere from 44% to 50% with 95% confidence.

What the margin of error does not capture: systematic error. The MOE only quantifies random sampling variation — the luck of which specific people happened to answer the phone. It says nothing about whether the sample was unrepresentative, whether the likely voter model was wrong, whether certain groups were underweighted, or whether the question wording pushed respondents toward a particular answer. These non-random errors — which are far more consequential in modern polling — are not reflected in the margin of error at all. This is why polls can show systematic errors of 5-7 points in one direction across an entire cycle while reporting a margin of error of ±3%.