In an ideal world, researchers might have enough time, money and research assistants to include census-like studies of entire populations in their research designs. In reality, this is entirely implausible, but luckily statistics implies that this is also entirely unnecessary. Sampling is much like the logic of flipping a coin – if you only flip it six times, the chances of getting an accurate 50/50 distribution is diminished, but if you flip it one hundred times you’ll probably get something a lot closer to 50/50. In the same way, analyzing an entire population only requires a researcher to study a large enough chunk to ensure that their analysis is not significantly influenced by random chance and errors.
Generalizability is often considered to be a paramount goal in scientific inquiry, although this is not entirely the case in social sciences. Generally speaking, generalizability is a typical goal of quantitative analysis in the social sciences, but somewhat less so for some instances of qualitative analysis. A typical line of reasoning for not striving for generalizability might be that access to the population that is being studied is limited and a representative sample could never be selected; researchers with constructivist viewpoints might also conclude that generalizability is an impossible goal to achieve. That said, even when generalizability is not a plausible goal, the specific sampling technique used is still important.
When generalizability is a plausible goal, the sampling techniques used are termed as “probability sampling methods.” By utilizing probability mechanics, these sampling methods avoid systematic biases caused by underrepresentation and overrepresentation. The most obvious of these methods is simple random sampling, where subjects are selected with a random numerical component to ensure that selection is entirely random. Closely related to this sampling technique is systematic random sampling, where specific items in sequential files are selected according to the necessary representative sample size and the total number of available entries. If it is essential for a researcher to make sure their sample accurately represents a specific distribution of a characteristic in a population, proportional stratified sampling is used; similarly, if a researcher wants to focus on a specific (or underrepresented) characteristic in a population, disproportionate stratified sampling, in combination with weighting techniques, is used. Finally, researchers can use cluster sampling as an alternative to simple random and systematic random sampling techniques. Cluster sampling focuses on random selection of naturally clustering elements of a population (such as city blocks or schools), and while it is sometimes convenient for survey researchers, it is unfortunately more likely to cause sampling errors.
When generalizability is not a plausible goal, nonprobability sampling is used. The most obvious type of these sampling techniques is availability sampling, where the sample is simply selected by convenience (such as people passing by on the street). The use of this technique varies contextually, but given the difficulty of ensuring that all subjects being studied belong to a particular population, it is not often applicable to nuanced research questions. More preferable techniques include quota sampling and purposive sampling. Quota sampling represents an attempt to mitigate the detrimental effect of nonprobability based sampling methods – researchers attempt to meet sampling quotas that represent a population, but since all possible characteristics of a population can’t realistically be identified, this sampling technique still fails to lead to a generalizable sample. Purposive sampling techniques are used when the sample is selected to fulfill a specific research purpose, such as only interviewing people who are particularly knowledgeable about a subject or the study of a small subset of a population. Finally, there is snowball sampling, where each subject provides the researcher with a list of other cooperative subjects – this is best utilized when studying populations with limited access, such as police officers.