Polls are reported as measurements and are estimates with error, produced by processes that involve substantial judgement.
Sampling
A sample intended to represent a population.
Which is the entire foundation, and random sampling has become extremely difficult as response rates have collapsed.
Response rates in the low single digits are common for telephone polling.
Weighting
Adjusting the sample to match known population characteristics.
Which corrects for who responded and introduces assumptions about who did not.
Weighting choices — which variables, and how — differ between pollsters and materially affect results.
Margin of error
Reflects sampling variation only.
Which means it understates total uncertainty, since it excludes weighting decisions, question wording and non-response bias.
The stated figure is a floor rather than a full accounting.
Likely voter models
Estimating who will actually vote.
Which is a substantial modelling exercise and a major source of divergence between polls.
Different assumptions about turnout produce different headline numbers from identical raw data.
House effects
Systematic differences between pollsters arising from methodology.
Which are measurable and reasonably stable.
Averages across pollsters address this partially.
Question wording and order
Both affect responses measurably.
Which is why reputable pollsters publish full question wording.
A poll not publishing its questions is providing less information than it appears to.
Historical accuracy
Polling has performed well in many elections and has missed substantially in several.
Which prompted methodological reviews and changes in weighting practice.
Errors have tended to be correlated across pollsters rather than random.
Reading polls sensibly
Look at averages, look at trends rather than single results, and treat small leads as uncertain.
Published methodology statements are the way to distinguish serious polling from the rest.
Online panels
Most polling now uses recruited panels rather than random telephone samples.
Which is cheaper and produces a sample that is not random by construction.
Weighting and panel management are what make these usable, and both involve judgement.
Modelling versus polling
Forecasts combine polls with historical patterns, economic data and constituency-level modelling.
Which is a different exercise from measuring current opinion.
Forecast probabilities are frequently misread as predictions of certainty.
Exit polls
Conducted with voters leaving polling places, using established methodology.
Which has generally been more accurate than pre-election polling.
Postal and early voting complicate exit polling and require separate adjustment.
Reporting standards
Polling associations publish standards on disclosure of methodology, sample and dates.
Which reputable pollsters follow.
Polls not disclosing these should be treated with corresponding caution.
Reading a poll
Fieldwork dates, sample size, method, weighting and the exact question.
Which are published in tables and almost never appear in coverage.
Undecided voters
How pollsters handle people who will not state a preference varies substantially.
Which includes excluding them, allocating them proportionally or by past behaviour, or reporting them separately.
This choice alone can shift a headline figure by several points.
Shy voter effects
The hypothesis that some respondents misreport their intention.
Which has been proposed after several polling misses and is difficult to demonstrate.
Differential non-response — some groups being less likely to answer at all — is a better-supported explanation for the same errors.
Constituency and state-level polling
Smaller samples with larger errors than national polls.
Which matters enormously in systems where geographic distribution determines outcomes.
Multilevel modelling using large national samples to estimate local results has become a standard alternative.
Frequency and herding
Pollsters publishing near an election may be reluctant to release outlying results.
Which produces artificial convergence and has been documented statistically.
Genuine variation should produce more spread than is sometimes observed.
The reasonable position
Polls measure current stated intention with meaningful uncertainty, and treating them as forecasts is where most misinterpretation begins.
What polls are actually good for
Measuring the direction and rough magnitude of movement in opinion over time.
Which they do reasonably well when read as a series rather than as individual results.
Issue polling — what people say they care about — is frequently more informative than voting intention and is reported less.
The persistent misreading
Treating a two-point lead as a fact rather than as a number with uncertainty around it substantially larger than two points.
Which accounts for most of the surprise when results differ from expectations.
A closing observation
Polling has become harder as response rates collapsed, and the methodological work required to produce usable estimates from unrepresentative samples has grown correspondingly.
That work is legitimate and involves judgement, which is why two competent pollsters can produce different numbers from the same electorate on the same day.
Neither is necessarily wrong.
Polling in other contexts
Referendums, primaries and low-turnout elections are harder to poll than general elections.
Which is because turnout modelling is more uncertain and historical patterns are less applicable.
Errors have been larger in these contexts, consistently.
What to read
Published tables including fieldwork dates, sample composition, weighting variables and full question wording.
Which reputable pollsters release alongside every poll and which coverage almost never references.