Snowball Sampling

Snowball sampling is a nonprobability sampling method in which existing study participants recruit additional participants from their social networks, creating a chain of referrals that expands the sample.^[3] It is particularly useful for reaching hidden, stigmatized, or hard to access populations that cannot be identified through conventional sampling frames.^[1][3]

Some research populations are not enumerable through standard sampling frames. Individuals who use illicit drugs, undocumented migrants, people with stigmatized health conditions, and members of small subcultures may be inaccessible through population registries or random sampling methods. Snowball sampling provides a practical way to recruit these populations by using trust and social connections.^[1][3] While it does not produce statistically representative samples, it enables research that would otherwise be impossible. Because people tend to refer others like themselves, and well connected people are more likely to be reached, estimates from a snowball sample can be biased in ways that a larger sample does not fix, so findings need to be framed as describing the people reached.

The researcher begins by identifying and recruiting a small number of initial participants (seeds) who meet the study's eligibility criteria. These seeds are asked to refer other eligible individuals from their social networks. The referred individuals, once enrolled, are asked to refer additional participants, and so on.^[3] The process continues until the desired sample size is reached or until referral chains are exhausted.

The selection of initial seeds significantly affects the sample composition. Seeds who are well connected and trusted within the target community tend to produce longer and more productive referral chains. Using multiple seeds from different subgroups within the population helps broaden the sample's coverage.

A variant called respondent driven sampling (RDS), introduced by Heckathorn in 1997 and refined in later work, structures the referral process: each participant receives a small number of coded coupons to pass to peers, participants are rewarded both for taking part and for recruiting eligible peers, who recruited whom is recorded,^[1] and every participant reports their network size (degree). RDS estimators then weight participants to adjust for differences in network size and for the tendency of individuals to refer people similar to themselves (homophily).^[4] These estimates rest on strong assumptions, for example that participants recruit at random from their personal networks, that ties are reciprocal, that network sizes are reported accurately, and that referral chains run long enough for the sample to stop depending on the seeds.^[2] When the assumptions fail, RDS estimates can be biased, and they are often much less precise than a simple random sample of the same size would be.^[5] RDS has been used extensively in HIV surveillance among key populations.

Use snowball sampling when the target population is difficult to identify or access through conventional sampling methods, when trust and personal connections are necessary for recruitment, and when no sampling frame exists. It is appropriate for exploratory and qualitative research, pilot studies, and, in its respondent driven form, surveillance of hidden populations. Do not use snowball sampling when a representative sample is required for statistical generalization, as the nonprobability nature of the method limits inference to the broader population.

A study of hepatitis C prevalence among people who inject drugs in a mid sized city uses snowball sampling because no registry of this population exists. The research team recruits 8 initial seeds from a harm reduction clinic and a needle exchange program, representing different age groups and neighborhoods. Each seed refers 2 to 5 contacts who meet the eligibility criteria. After four referral waves, the study enrolls 186 participants. Hepatitis C antibody prevalence in the sample is 54% (100 of 186), which the researchers report as a sample figure rather than a population estimate. The researchers note that the sample may overrepresent individuals connected to harm reduction services and acknowledge this limitation in discussing generalizability.