Sample

A sample is a subset of a larger group, drawn from it and measured so that conclusions about the whole group can be inferred without examining every member. The larger group is the population, all the individuals or observations you actually care about, and the sample is the smaller slice you take from it. Sampling exists because measuring an entire population is often impossible, too slow, or too expensive. You cannot survey every customer, test an ad on every prospect, or observe every transaction, so you take a sample, measure it, and use what you find to estimate what is true of the whole. The entire logic rests on one condition. The sample must be representative, meaning its makeup mirrors the population well enough that what holds in the sample is likely to hold in the population. A sample that fails this condition produces confident conclusions that are simply wrong.

Samples matter because nearly all practical measurement in marketing and research is sampling. Surveys question a sample of customers, A/B tests expose a sample of visitors to each variant, and market research infers the preferences of a whole market from a sample of respondents. In every case, the point is to learn about the population by studying a manageable subset. This is powerful, since a well-chosen sample can describe a vast population accurately at a fraction of the cost of measuring everyone. But the power depends entirely on the quality of the sample. If the way the sample was drawn systematically over- or under-represents parts of the population, the results are biased, and the estimates drawn from them mislead. Two things govern whether a sample can be trusted, how it was selected and how large it is.

A sample and a population are a matched pair, and keeping them straight is the foundation of sampling. The population is the complete group you want to understand, every customer, every visitor, every unit. The sample is the subset you actually measure. A statistic computed on the sample, such as a sample mean or sample conversion rate, is an estimate of the corresponding figure in the population. The whole exercise is inference, using the observed sample to reason about the unobserved population. The gap between what a sample shows and what is true of the population comes from two sources. Sampling error is the ordinary random variation that shrinks as the sample grows, and bias is systematic distortion from how the sample was chosen, which a larger sample does not fix.

Two properties decide whether a sample supports sound conclusions, representativeness and size. Representativeness is about how the sample was selected. A random sample, where every member of the population has a fair chance of being included, tends to be representative, while a convenience sample of whoever was easiest to reach often is not, baking in bias that no amount of analysis removes. Size is about precision. A larger sample reduces random sampling error and narrows the uncertainty around estimates, but size cannot rescue a biased sample, since a big unrepresentative sample is just confidently wrong. The two properties are distinct and both necessary. A sample must be selected without systematic bias to be trustworthy at all, and it must be large enough to be precise enough for the decision at hand. Get either wrong and the inference fails.

Draw samples in a way that gives every relevant part of the population a fair chance of inclusion, favoring random or otherwise principled selection over convenience, because representativeness is what makes inference valid in the first place. Size the sample to the precision the decision needs, since a larger sample narrows random error, but never lean on size to compensate for a biased draw, because it cannot. Be explicit about what population the sample is meant to represent, and check that the sampling method actually reaches it rather than a skewed slice of it. When you report a result from a sample, treat it as an estimate with uncertainty, not a fact about the whole population, and let the sample's quality and size temper how firmly you act on it.

The failures divide into problems of selection and problems of size. A biased sample, drawn in a way that over- or under-represents parts of the population, yields skewed conclusions that no analysis can correct, and this is the more dangerous error because a large biased sample looks authoritative while being wrong. A sample that is too small carries so much random variation that its estimates are unreliable and can flip from one draw to the next. Generalizing from a sample to a population it never represented, such as inferring the whole market from a self-selected group of the most engaged customers, is a version of the bias trap. The disciplined approach selects the sample without systematic bias, sizes it for adequate precision, and reads every sample statistic as an uncertain estimate of the population.

A sample is a subset drawn from a larger population and measured to infer facts about the whole, valid only when the sample is representative and adequately sized.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where sample is a core concern: