Sampling Bias
A good sample is unbiased, meaning that the sample has the same attributes as the population of interest.
In other words, if a sample does not accurately represent the population of interest, then the sample is biased.
Different types of sampling biases
Non-random sampling bias
- The sample must be randomly extracted from the population - You cannot cherry-pick!
- You should not perform convenience sampling!
Voluntary response bias
- Members decide if they if want to be part of the sample or not.
- This biased sample will not be representative of the whole population.
- Example People who are bored or have very strong opinions about a particular subject are more willing to participate (i.e. self-selection). You will end up with a biased sample full of bored and very opinionated people.
Nonresponse bias
- If your sampling method is prone to low response (e.g. mailing surveys) as individuals that are chosen may be unable or unwilling to participate.
Undercoverage bias
- This happens when a specific type of individuals in a population is not adequately represented in the sample.
- Example Females make up only 29% of engineering majors at UCLA. If you only survey students near the men’s washroom at the Engineering Student Centre (i.e. convenience sampling), then females may likely be unrepresented in your sample.
Recipe for a Good Sample
When determining the best sampling method, we need to consider the sample size, randomization of the sample, and potential sampling biases.
Size
- A sample needs to be large enough according to the Central Limit Theorem but it does not need to be humungous.
- There is no magic fraction of the population that we need to sample, but generally we do not sample over 10% of the population.
- Additionally, we have no control over the population size.
Randomization
- Obtaining a random sample is key.
- Randomly selecting subjects in a sample minimizes the chance of obtaining a biased and non-representative sample.
Bias
- Your sample needs to be free from bias.
- A biased sample is not representative of the population of interest.