Four ways to choose who gets measured
Sampling method is where studies are won or lost: the analysis can be flawless and the sample still wrong. Four designs cover most honest work. Each trades cost, effort and precision differently, and each fails in its own characteristic way.
Simple random sampling
Every member of the population has an equal chance, drawn by lottery or random numbers from a complete list. It is the benchmark the mathematics assumes. Its weakness is logistics: you need the list (a sampling frame), and pure chance can still hand you a lopsided draw from a small population. Example: number the bakery's 1,200 weekly loaves by batch and position, draw 60 numbers, weigh those loaves.
Stratified sampling
Split the population into strata that differ on the thing you measure (undergraduates and postgraduates; weekday and weekend loaves), sample randomly inside each, and combine with the strata's true shares as weights. Proportional allocation mirrors the population: a college of 700 undergraduates and 500 postgraduates contributes 70 and 50 to a sample of 120. Stratification removes the risk of a stratum going missing by luck and usually buys precision for free.
Systematic sampling
Pick a random start, then take every kth member: for 60 from 1,200, every 20th loaf. Simple to run on a production line or a list. The trap is periodicity: if the list itself cycles with a pattern that matches k (every 20th item is always the first off a mould), the sample silently becomes a sample of one position. Shuffle or verify the list's order first.
Cluster sampling
Randomly pick whole groups (schools, shops, villages) and measure everyone inside the chosen ones. It slashes travel and cost, which is why national surveys use it. The price is precision: members of a cluster resemble each other, so 500 people in 5 schools carry less information than 500 people drawn individually. Effective sample size, not headcount, is the real currency.
Convenience sampling wearing a lab coat: the first 100 customers, students in your own class, followers who answered the poll. None of the four designs above is approximated by grabbing who is near. If you must use a convenience sample, describe findings as being about those respondents, and let the reader decide how far they travel.
Nonresponse: the sample you planned vs the one you got
Every design degrades the same way: selected people who cannot or will not answer often differ from those who do. Report the response rate, compare respondents with the frame on what you know (age bands, regions), and treat a low response rate as a finding about your method, not a footnote.
Frequently asked questions
What is the difference between stratified and cluster sampling?
Stratified samples from every group to guarantee coverage; cluster samples whole groups and skips the rest to save cost. Strata should be internally similar but all represented; clusters are a convenience you randomize over.
How large must a sample be?
It depends on the precision you need: the survey sample size calculator works it out from your margin of error and confidence level. For proportions at 95% confidence and a 5-point margin, the answer is 385 before any finite-population relief.
Why not just sample everyone reachable online?
Reachability is a selection mechanism: it excludes the offline, the busy and the indifferent, and overcounts the motivated. That bias does not shrink with sample size; it sharpens.