Sampling bias is a problem in how a sample is chosen, not in how the data is analysed afterwards.
It makes some members of the population more likely to be included, so the sample leans in one direction.
No method applied later can fully remove it, because the people who were missed left no data.
Undercoverage happens when part of the population is missing from the list, such as people without internet access in an online survey.
Non-response bias happens when the people who decline to take part differ from those who agree.
Voluntary response bias happens when people choose to join, so the most motivated dominate.
Survivorship bias happens when only the survivors of some process are visible, such as studying only companies still trading.
For example, studying only successful start-ups makes risky strategies look safer than they are.
Convenience samples, such as surveying the people in one office, are quick but rarely represent anyone else.
Bias does not shrink with sample size, unlike random error, which falls as the sample grows.
A huge biased sample is dangerous because its apparent precision makes a wrong answer look trustworthy.
Checking the sample against known population figures, such as age or region, reveals some bias but not all.
Weighting the sample to match known figures can reduce bias in those variables, but not in ones nobody measured.
In practice, describing exactly who was asked and who answered is the best protection for readers.
The common misconception is that a large sample cannot be biased, when size and bias are separate problems.