What sampling is: types and examples

Sampling is the technique of studying a part of a group to draw conclusions about the whole group. It's used because asking everyone, or measuring everything, is usually impossible, expensive or unnecessary. An election poll, tasting a batch of wine or the poll you post to your followers in stories are all samples.

The word has a second, very common use in audio and video: taking "samples" of a continuous signal to turn it into digital. We explain it at the end.

The basics

• Population: the whole group you want to conclude something about. For example, everyone who follows your account.
• Sample: the part you actually study. For example, the 300 who answer the poll.
• Representativeness: how closely the sample resembles the population.
• Bias: a systematic error that makes the sample unlike the population.
• Margin of error: how far the sample's result may be from the whole population's.

Types of sampling

Probability sampling

Every member of the population has a known chance of being in the sample. It's what lets you generalise rigorously.

• Simple random: everyone has the same chance, like drawing slips from a hat. Example: picking 200 customers at random from the full list.
• Systematic: you pick one in every k. Example: one in every ten orders to check shipping quality.
• Stratified: you split the population into groups (age, country, customer type) and sample within each in proportion to its size. Useful when groups are very different.
• Cluster: you pick whole groups at random instead of individuals. Example: choosing ten shops and surveying all their customers on one day.

Non-probability sampling

Selection isn't random. It's faster and cheaper, but the results can't simply be extended to the whole group.

• Convenience: you ask whoever is at hand. A stories poll is the perfect example: those who see it and feel like it answer, not your followers in general.
• Quota: you set how many people of each type you need and fill them up. Better than pure convenience, but within each quota there is still no randomness.
• Snowball: each participant brings the next. Useful for hard-to-reach groups, such as professionals in a niche.
• Purposive or judgement: you choose cases you consider representative. Good for in-depth interviews, not for figures.

How many responses you need

As a guide for large populations:

• To discover problems or ideas: 5 to 10 in-depth interviews usually surface the main patterns.
• To see general trends: between 100 and 400 responses.
• For a figure with a 5% margin of error and 95% confidence: about 385 responses, as long as the sample is random.

If you will compare subgroups (for example men and women, or Spain and Mexico), each subgroup needs its own minimum. And a large but biased sample is still misleading: a thousand answers from your most active fans don't speak for your whole audience.

Biases that ruin a sample

• Selection: the way you recruit leaves part of the population out. An online survey doesn't reach people who aren't online.
• Non-response: those who don't answer differ from those who do. Unhappy customers tend to ignore the satisfaction survey.
• Question wording: a question that pushes towards an answer ("Did you love our new design?").
• Social desirability: people answer what sounds good, not what they do.

If you use polls on social media, the guide to Instagram polls explains how to word questions that reduce these biases.

Sampling in audio and video

In digital audio, sampling means measuring the sound signal many times per second and storing each measurement as a number. How many times per second is the sample rate, in hertz:

• 44.1 kHz (44,100 samples per second): the CD standard and common in music.
• 48 kHz: the standard in video, film and TV. It's the one to use when recording and exporting videos.
• 96 kHz or higher: professional recording when the sound will be heavily processed.

The rule behind it is the Nyquist-Shannon theorem: to reproduce a frequency properly you must sample at least twice as fast. Since human hearing reaches about 20 kHz, 44.1 or 48 kHz is enough for listening.

A classic problem when editing video is mixing 44.1 kHz audio into a 48 kHz project: depending on the software, it can cause slight drift or quality loss on conversion. The safest option is to record everything at 48 kHz.

In images there is also chroma subsampling (4:4:4, 4:2:2, 4:2:0): colour is stored at a lower resolution than brightness to save space. Most phone and social media videos use 4:2:0.

Frequently asked questions

What is sampling in statistics?

It's the process of selecting part of a population (the sample) to study it and estimate what the whole population is like. The more representative the sample and the fewer biases it has, the more reliable the conclusions.

What are the types of sampling?

They split into probability types (simple random, systematic, stratified and cluster), where everyone has a known chance of being chosen, and non-probability types (convenience, quota, snowball and purposive), which are faster but less generalisable.

What's the difference between probability and non-probability sampling?

In probability sampling selection is random and the margin of error can be calculated, so results can be extended to the population. In non-probability sampling selection depends on availability or on the researcher's judgement, and results only describe those who answered.

What sample rate should I use for video?

48 kHz. It's the standard for video, TV and platforms such as YouTube. 44.1 kHz is the music CD standard; if you mix both in one project, convert everything to 48 kHz before editing.