Analysing every record isn't always possible or necessary. A well-drawn sample can give accurate answers faster and cheaper.
Simple Random Sampling
Every member of the population has an equal chance of selection. The baseline method — easy to analyse, but may under-represent small groups.
Stratified Sampling
Divide the population into groups (strata) — regions, customer tiers — and sample within each. Guarantees representation of every group and often improves precision. Sample small but important groups more heavily and weight results accordingly.
Cluster Sampling
Randomly select whole groups — schools, stores — and study everyone (or a sample) within them. Cheaper for field work, but less precise, because members of a cluster tend to be similar.
Systematic Sampling
Select every k-th record from a list. Simple, but dangerous if the list has a periodic pattern that lines up with k.
Sources of Bias
- Selection bias: the sampling method excludes part of the population (online surveys miss people who aren't online).
- Non-response bias: people who respond differ from those who don't.
- Survivorship bias: only surviving cases are observed.
- Convenience samples: whoever is easiest to reach.
Sample Size
Larger samples reduce random error but not bias. A huge biased sample is still wrong. Calculate the size needed for the precision you require.
In Data Work
Sampling large datasets for exploration is fine; check that the sample preserves important patterns, and use the full data for final figures when feasible.