Polls and Surveys

How Polls Work

A well-designed poll surveys a representative sample of a population — typically 1,000 to 1,500 people for a national poll. If the sample is truly random, this is enough to estimate the views of millions with a margin of error around ±3 percentage points.

The key challenge is representativeness. Pollsters must ensure their sample mirrors the actual population in terms of age, gender, race, education, geography, and other factors. When certain groups are harder to reach — young people, rural residents, non-English speakers — pollsters apply statistical weights to adjust for under-representation.

Why Polls Sometimes Fail

Polling failures have become more visible. The 2016 US election and the 2016 Brexit referendum both produced results that surprised many pollsters. Several factors explain polling errors:

Non-response bias: People who agree to take polls may systematically differ from those who do not. Response rates for phone polls have fallen from 36% in 1997 to under 6% today.

Social desirability bias: People may be reluctant to express politically unpopular views, leading to underestimates of support for controversial candidates or positions.

Likely voter models: Polls must guess who will actually vote. Small errors in turnout assumptions — especially among irregular voters — can produce large errors in predicted outcomes.

Worked Example: Margin of Error and the 'Statistical Tie'

A poll is a sample, so it carries sampling error. For a simple random sample the 95% margin of error is roughly 1 divided by the square root of the sample size. That single relationship explains most of what confuses people about polls:

Two lessons jump out. First, precision improves with the square root of the sample, so quadrupling the sample only halves the error — going from 1,000 to 2,000 respondents barely helps. Second, a modest national sample of about 1,000 people already achieves roughly +/- 3 points, which is why serious polls rarely need to be huge.

Now apply it. A poll of 1,000 voters shows Candidate A on 48% and Candidate B on 45%. The 3-point lead looks decisive, but with a margin of about +/- 3 points each, A's true support could be anywhere from 45% to 51% and B's from 42% to 48%. The intervals overlap, so this is a statistical tie — the data cannot confidently say who is ahead. Reporters who trumpet a 'commanding lead' from a within-margin gap are misreading their own numbers.

Crucially, the margin of error captures only random sampling error. It says nothing about bias from a skewed sample, bad question wording, or people who refuse to answer. A poll can be perfectly precise and still systematically wrong.

Read the Fine Print

Before trusting any poll, hunt for four details usually buried at the bottom. Who was sampled (registered voters, likely voters, or all adults — these can differ by several points)? How were they reached (live phone, online panel, text), since each method reaches different people? When was it in the field, because opinion moves after major news? And what was the exact question wording? A poll that hides its methodology deserves less trust than one reporting a slightly worse headline number transparently. The margin of error is the least of your worries if the sample or questions are skewed.