The idea behind the plus-or-minus
Imagine a bowl of marbles, some red and some blue, and suppose you want to know the share that are red without counting them all. You scoop out a handful and count. The share in your handful will usually be close to the share in the bowl, but not exactly the same. Scoop again and you will probably get a slightly different share.
That scatter between one scoop and the next is sampling uncertainty. A poll is a scoop from a very large bowl of people. The margin of error is a way of putting a rough size on how far a typical scoop might sit from the true value, purely because of the luck of who was drawn.
What the number means
When a poll reports a figure such as 50 per cent with a margin of error of plus or minus three points, the intended reading is that the true value in the population is probably somewhere between 47 and 53 per cent. “Probably” is doing real work here. The usual convention is a 95 per cent confidence level, which means that if the same poll were repeated many times with fresh random samples, the calculated ranges would capture the true value in roughly 95 of every 100 repetitions. It does not mean there is a guarantee, and in some repetitions the range will miss.
The margin is typically larger for smaller samples. As a rough rule, a random sample of around a thousand people often gives a margin of about three points, while a sample of a few hundred gives a noticeably wider one. Collecting more responses narrows the margin, but with diminishing returns: to cut it in half you generally need about four times as many respondents.
Reading gaps between two numbers
The margin of error applies to each figure, not to the gap between them. Suppose a poll shows one option at 48 per cent and another at 46 per cent. Each figure has its own uncertainty, and the difference between them is more uncertain than either one alone, often by a notable amount. A two-point lead in a poll with a margin of around three points is generally too small to distinguish from a tie.
This is why careful commentators describe such results as “no clear lead” rather than naming a leader. It is also why a change of a point or two between successive polls is often better treated as noise than as news.
Subgroups are less certain
A poll’s headline margin refers to the whole sample. When results are broken down by age group, region or any other category, each subgroup contains fewer people, and the margin around its figure is wider, sometimes much wider. A claim that young adults in one region have moved sharply may rest on only a few dozen respondents. Pollsters sometimes publish subgroup margins, and when they do not, it is reasonable to be more cautious.
What the margin of error does not cover
This is the point most often missed. The margin of error describes only random sampling variation, and it assumes the sample was drawn in a way that supports that calculation. It typically does not account for:
- Non-response. If certain kinds of people systematically decline to answer, the sample can lean one way no matter how large it is. That problem is explored under non-response bias.
- Coverage gaps. People missing from the list used to draw the sample have no chance of being included.
- Question wording and order. Leading or confusing questions shift answers in ways unrelated to sample size.
- Dishonest or inaccurate answers. Respondents may not report what they truly think.
- Adjustments. Statistical corrections such as survey weighting can add uncertainty, and some pollsters report a larger “design effect” margin to reflect that.
So a poll with a small margin of error is not automatically accurate. It may be precisely wrong. Precision (how tightly estimates cluster) and accuracy (whether they cluster around the truth) are different things.
Margins and opt-in samples
A formal margin of error rests on the idea that people were selected at random from a defined population. Many online surveys use volunteers who chose to join a panel. Some publishers still print a margin for these, and others describe an equivalent figure, often as a “credibility interval”, based on modelling assumptions. Those numbers can be useful, but they depend on the model being right, which is a different kind of claim from the simple scoop-from-a-bowl logic. The distinction is a main theme of the comparison between opt-in polls and random samples.
A practical way to use it
When you see a poll result, a few quick habits help:
- Find the sample size and the stated margin, and note whether it is for the whole sample or a subgroup.
- Compare any gap to the margin before treating it as a lead or a shift.
- Look at several polls. If they cluster, the picture is firmer than any single result suggests.
- Remember that the margin speaks only to chance, and ask separately whether the poll was well designed.
The margin of error is best seen as a modest, honest admission of uncertainty. It is a useful reminder that a poll is an estimate, and that small movements in headline numbers often say less than the news cycle implies.