Snowball sampling is a non-probability method in which the first participants, called seeds, refer or recruit other people from the same population, who then refer others, wave after wave. Researchers use it when no list of the population exists, as with people who inject drugs or undocumented migrants. The sample follows social ties, so whoever the seeds know shapes the results.
Key Takeaways
- A snowball sample, also called a chain-referral sample, grows as the seeds name or recruit peers, and then each new participant does the same.
- Leo Goodman’s 1961 formal version started from a random sample and measured relationships in a population; today’s version starts from whoever is reachable.
- Its biases are seed dependence, homophily (people refer people like themselves), volunteerism, masking (protecting friends by not referring them) and over-representation of well-connected people.
- In our simulation, five seeds from one community put a trait that affects 19% of the population at 10.5%, and longer chains did not move the estimate.
- Respondent-driven sampling (RDS) adds coupons, recruitment caps and weighting by network size, but in a Winnipeg study, two arms started from different seeds gave different estimates.
What is snowball sampling?
A snowball sample starts with a few people from the target group and grows by asking each participant to refer or recruit others who meet the criteria. It stops when the researcher reaches a target size or, in qualitative work, when new interviews stop adding new information. Nobody’s chance of selection is known, so it is a non-probability method.
Douglas Heckathorn describes how the method is actually used: “ideally, a randomly chosen sample serves as initial contacts, though in practice ease of access virtually always determines the initial sample.”
The method exists for what Heckathorn calls hidden populations, which have two features: “no sampling frame exists, so the size and boundaries of the population are unknown,” and membership is stigmatised or illegal, so people refuse to take part or give unreliable answers.
Goodman’s original design was different. The best-known formal definition comes from a 1961 paper by the University of Chicago statistician Leo Goodman, who credited the sociologist James Coleman with introducing him to the topic. It begins: “A random sample of individuals is drawn from a given finite population.” Each person in it names k others (his examples include “k best friends”), the people named form the first stage, they name k others in turn, and so on for s stages. In most of the paper, every person entered the first sample with a known probability p, decided independently, like a coin toss. The aim was to estimate features of the network itself, such as the number of pairs who would name each other. For that count, one stage of naming, with the people named interviewed too, beat a random sample in which only the sampled people were interviewed: the plain sample’s estimator had at least twice the variance, even when the plain sample was expected to be as large as the snowball design at its largest. Today’s method keeps Goodman’s referral chains and drops his random start.
Snowball, chain referral and respondent-driven sampling: the main variants
The methods literature describes three variants. Goodman’s design starts from a random sample and studies the network. Chain-referral sampling, the everyday version, starts from whoever the researcher can reach and uses referrals to find more participants. Respondent-driven sampling adds recruitment coupons, limits and weighting so that the final estimates can be adjusted for how people were recruited.
| Variant | How the seeds are chosen | How new people enter | What it is built for | What you can claim |
|---|---|---|---|---|
| Goodman’s snowball design (1961) | A random sample, each person included with a known probability | Each person names k others, for s stages | Estimating network features, such as the number of mutual pairs | Statistical estimates about the network |
| Chain-referral sampling (everyday use) | Whoever the researcher can reach: contacts, clinics, organisations | Participants name people or pass on the study details | Finding members of a group with no list; qualitative interviews | Findings about the people reached, not population figures |
| Respondent-driven sampling (Heckathorn, 1997) | A handful of seeds, ideally varied | Participants hand out a fixed number of coupons; recruiters are paid per recruit | Prevalence estimates in hidden populations | Population estimates, if its assumptions hold |
Biernacki and Waldorf described the second row in a 1981 article whose abstract notes that chain referral “has been widely used in qualitative sociological research, especially in the study of deviant behavior,” yet its procedures “are not self-evident or obvious,” and the article draws on their exploratory study of ex-opiate addicts.
Study guides often split snowball designs into linear, exponential non-discriminative and exponential discriminative types. These labels do not appear in Goodman’s, Heckathorn’s or Salganik and Heckathorn’s papers, so this guide sticks to the variants in the table.
How does the snowball method work, step by step?
The researcher recruits a few seeds who belong to the target group, interviews them, and asks each one to name or recruit others who fit the criteria. Those recruits are interviewed and asked the same, and each round is a wave. Recruitment stops at a planned size or at saturation. The steps are simple; the decisions inside them shape who ends up in the sample.
- Define who qualifies in terms you can check. Heckathorn’s 1997 Connecticut study confirmed drug injection through a seven-step screening protocol, because paying people to recruit invites recruits who do not belong.
- Choose the seeds: how many, where you find them and how much they differ. The simulation below shows how much rides on this.
- Decide how referrals happen. Participants either give you names, or pass study details to peers who then contact you, which protects people who do not want to be named. A cap on referrals per person stops a few well-connected people from filling the sample.
- Record the chain and stop on a rule written in advance: who referred whom, which wave each person came from, and a target size or saturation.
Snowball sampling examples from real studies
Published examples cover people who inject drugs, sub-Saharan migrants in Morocco (most of them undocumented) and patients with rare diseases: groups with no list to sample from and, in the first two cases, good reasons to stay out of sight. Each study below reports how the first participants were found, how referrals worked and how many people took part.
People who inject drugs (Connecticut, 1994–1996). Heckathorn’s respondent-driven sample surveyed 277 active drug injectors in two small Connecticut cities. Every participant, seeds included, started with three coupons and was paid $10 for each peer who came for an interview. The chains grew large: at one site, a single subject generated more than one hundred recruits.
Sub-Saharan migrants (Morocco, 2008). Keygnaert and colleagues trained twelve sub-Saharan migrants as community researchers, who served as the first seeds and interviewed peers about violence, including sexual violence, at four sites. Each respondent received 100 dirhams (about €9) and a coupon worth 25 more for bringing in another eligible migrant, giving 3 to 4 waves per researcher. Of the 154 people interviewed, 107 (69.5%) were undocumented.
Rare-disease patients (Hong Kong, 2018). Chung and colleagues had no patient registry to sample from, so they combined purposive, opportunity and snowball recruitment through the patient group Rare Disease Hong Kong, with referrals from 24 organisations and units. They interviewed 176 patients between April and September 2018, which they estimate at about 0.16% of the city’s roughly 110,000 people with a rare disease.
Is snowball sampling biased?
Yes, in predictable ways. The people who enter depend on who the seeds were, on who knows whom, on who agrees to take part and on who gets protected from being named. A snowball sample over-represents well-connected people and the communities around the seeds, and the usual sampling-error formulas do not measure any of this.
Heckathorn summarises Bonnie Erickson’s 1979 list of problems with chain-referral samples. With homophily added, which Heckathorn measured in his own data, the list has five named biases:
- Seed dependence. In Erickson’s words, “inferences about individuals must rely mainly on the initial sample, since additional individuals found by tracing chains are never found randomly or even with known biases.” When the seeds are a convenience sample, the whole sample inherits it.
- Homophily. People know, and refer, people like themselves. At Heckathorn’s first site, non-Hispanic white recruiters brought in 70% white recruits on average, and non-Hispanic Black recruiters 56% Black recruits. Chains stay inside the community where they started, so a group with few ties to the seeds’ community stays under-represented.
- Volunteerism. Chain-referral samples lean toward people who agree to take part, and the problem grows when the seeds are volunteers, who are unusual in how willing they are.
- Masking. Participants protect friends “by not referring them,” which matters most when membership is illegal or stigmatised.
- Network size. Because referrals travel along ties, “subjects with larger personal networks will be oversampled, and relative isolates will be excluded.”
Real data show the same. Wylie and Jolly ran two recruitment arms side by side in Winnipeg in 2009, one from 22 seeds chosen by study staff, one from 118 people who heard about the study by word of mouth and asked to join. Both drew on the same network over the same months, yet their respondent-driven estimates differed (staff-seeded arm first): 6% against 13% for sex work, 16% against 5% for street-involved youth, 10% against 3% for men who have sex with men. For four variables, the two arms’ confidence intervals did not overlap or overlapped by only 0.01. The broader problem of samples that differ from their population is covered in our guide to sampling bias.
How much do the seeds matter? A simulation
In a simulated population of 2,000 people where a trait affects 19%, chain-referral samples of 200 started from five seeds in one community put the trait at 10.5% on average. Seeds in the other community gave 39.4%. The two communities shared only 30 ties, so the chains rarely crossed, and the seeds’ community set the answer.
The population (constructed for this simulation). Community A has 1,400 people, 10% of whom have the trait. Community B has 600 people, 40% of whom have it. True prevalence is (140 + 240) ÷ 2,000 = 19.0%. Inside each community, every pair of people is linked at random with a chance that gives each person about six ties. Between the communities there are only 30 ties, the bridges.
The recruitment rule. Each sample starts from five seeds. Every participant gets three coupons and passes each to a different contact not yet in the sample. Everyone recruited agrees, so there is no refusal and no masking: a best case. Waves continue until 200 people are in, which took four waves in most runs. We ran 1,000 samples for each choice of seeds, in Python with a fixed random seed, and compared them with 1,000 simple random samples of 200.
| Seeds (five per sample) | Average share of the sample from community B | Average estimate | Middle 95% of estimates | Samples within 5 points of 19% |
|---|---|---|---|---|
| All five in community A | 1.0% | 10.5% | 7.0% to 15.0% | 6.6% |
| Three in A, two in B | 39.7% | 22.2% | 17.0% to 27.5% | 78.0% |
| All five in community B | 97.6% | 39.4% | 33.5% to 44.5% | 0.0% |
| Simple random sample of 200 | 30.0% | 19.0% | 14.0% to 24.0% | 95.9% |
Each set of seeds reproduced its own community: seeds in A gave an estimate close to A’s 10%, seeds in B one close to B’s 40%. Mixed seeds helped without solving the problem, because a 3-to-2 split put 40% of the sample in a community that holds 30% of the population. And a researcher with one all-A sample would see a tidy figure around 10% with nothing in the data to warn them.
Do more waves fix it? Not here. With seeds in A, the running estimate was 10.3% after the first wave (20 people), 10.2% after the second (62 people) and 10.5% after the third (185 people). The chains grew, but inside community A, the same pattern our guide to convenience sampling finds when a sample from one library gets bigger.
Do more bridges fix it? Partly. With 300 ties between the communities instead of 30, all-A seeds gave an average of 12.5%, with 8.2% of the sample from community B, and 30.8% of samples landed within 5 points of the truth. Better connected, but still far from the roughly 96% of random samples that landed within 5 points.
Can weighting rescue it? Only when both communities are in the sample in useful numbers. Weighting each community back to its known 70/30 share brought the mixed-seed samples to an average of 19.3% (middle 95%: 14.8% to 24.0%). With all-A seeds, 44.8% of samples held nobody from community B, and in the rest the middle 95% of weighted estimates ran from 5.3% to 38.8%, because B’s rate rested on a handful of people. Weighting also needs you to know which groups matter and how large they are, and it worked here partly because, inside each community, who knew whom had nothing to do with the trait.
How respondent-driven sampling tries to correct the bias
Respondent-driven sampling keeps the referral chains but changes how they run and how the data are analysed. Participants recruit peers with a limited number of coupons, recruiters are rewarded for each recruit, recruitment continues for many waves, and the analysis weights people by their network size. Under its assumptions, the final estimates no longer depend on who the seeds were.
Heckathorn introduced it in 1997 with four main features:
- A dual incentive. People were paid for their own interview and also for each peer they recruited, $10 per recruit in his study, at an average cost of about $14 per recruited subject.
- Coupons instead of names. Recruits contact the study themselves, which reduces masking: “a subject who would not feel comfortable giving a researcher the name of a peer, may nonetheless succeed in recruiting that peer.”
- A cap of three coupons per person, with three more only after three successful recruitments and a structured test showing those recruits had been taught about HIV prevention, to prevent “semi-professional recruiters” from dominating. The cap also pushed the sample through more waves.
- Running to equilibrium. If recruitment runs long enough, the sample’s make-up settles to a mix that does not depend on the seeds; for the recruitment patterns in his Table I, his calculation showed that getting within 2% of it never took more than six waves, whatever the seeds.
Salganik and Heckathorn (2004) added an estimator. Because chain-referral methods “overrepresent people with high degree,” meaning many contacts, it corrects for the number of contacts each person reports and for who recruited whom. Under the stated conditions, they show the resulting estimates are “asymptotically unbiased no matter how the seeds are selected.” Those conditions include a network that forms one connected component and recruiters who pick at random among their contacts. The authors doubt that recruiting over sparse “sexual or drug-sharing networks” is wise, and call the method “not appropriate for groups like tax evaders, who do not have frequent contact with each other.”
Our simulation shows why those conditions matter: with 30 bridges and four waves, the chains were nowhere near equilibrium. Wylie and Jolly, whose two Winnipeg arms both used three coupons but paid nothing for recruiting, concluded that seed selection “has the potential to influence which subgroups within a population are accessed.”
Snowball and purposive sampling in qualitative research
In qualitative research, snowball recruitment is a way to find information-rich participants, and the aim is depth, not population figures. Researchers often choose the seeds purposively, for characteristics the study needs, and then let referrals find more people like them. The sample is judged on whether it covers the experiences the question needs, not on whether it mirrors a population.
Naderifar and colleagues’ 2017 review calls the method purposeful in its title and “a convenience sampling method” in its background section, and says the first participants are “usually selected via convenience sampling.” Both labels fit part of the process: seeds may be picked for convenience or on purpose, while the referrals depend on who knows whom. In their account, recruitment continues until data saturation.
In practice, pick seeds that differ on what matters to the question, such as age, region or route into the group, and let referrals extend each seed’s reach: the qualitative version of mixed seeds. Our comparison of qualitative and quantitative research explains why the two judge samples differently, and the glossary entry on sample size covers the planning side.
How to report a snowball sample in your methods and limitations
Describe the seeds, the referral mechanism and the shape of the chains, because those details let a reader judge which networks the sample reached. Then, in the limitations, say which groups were likely over- or under-represented, which way that pushes your findings and who the results describe. Calling a snowball sample representative needs evidence from outside the sample.
In the methods, give the number of seeds, how you found them and how they differed; whether participants gave names or passed on study details, with any cap or incentive; the number of waves and how many seeds recruited nobody; and the stopping rule and final sample size. Our guide on writing a methodology section shows where these details sit. A model pair, with bracketed placeholders:
Methods: “Participants were recruited by chain referral from five seeds, three contacted through [organisation A] in [community A] and two through [organisation B] in [community B]. Each participant received three invitation cards to pass to people they knew who met the inclusion criteria, with no payment for recruiting. Recruitment ran for four waves, until 200 people had taken part.”
Limitations: “Because two of the five seeds came from [community B], 80 participants (40%) belonged to that community, against an estimated 30% of the population [source]. If [the outcome] is more common there, our overall figure overstates it. Participants with few contacts in either community are likely under-represented. The findings describe the networks reached from these seeds.”
Chung and colleagues’ rare-disease study shows a published version. They name the missing frame, the people missed (patients with severe disabilities who could not leave home) and both possible directions: “the potential sampling bias could go in both directions,” so the findings “primarily reflect the experiences of patients situated toward the middle of the distribution.” Our guide to the limitations of a study has more published examples and a template.
Where Papero fits
Papero is an AI writing workspace for students and researchers. It does not recruit participants, track referral chains or analyse your sample; it helps with the reading and writing around them.
Find papers searches more than 250 million papers through OpenAlex and saves any result to your source library, and Get free PDF downloads a saved paper’s open-access PDF when one exists. Chat answers questions about the PDFs in your library and the document you are writing, citing the passages it used, which helps when you compare how several studies chose their seeds or worded their limitations. In Review, the Claims check finds sentences that need a citation and offers papers to cite. What each plan includes is on the pricing page. The glossary entry on random sampling covers the probability side, and the thesis writing page shows how chapter outlines, one source library and version history fit together.
Frequently Asked Questions
What is non-probability sampling?
Non-probability sampling selects participants without giving every member of the population a known chance of selection. Convenience, purposive, quota and snowball samples are all non-probability methods. Because the selection chances are unknown, margin-of-error formulas do not show how far the sample is from the population, and findings apply to the people sampled.
What are the 5 basic sampling methods?
The five usually taught first are four probability methods, simple random, systematic, stratified and cluster sampling, plus convenience sampling, the most familiar non-probability method. The probability methods support population estimates with a known margin of error; convenience samples, like purposive, quota and snowball samples, describe the people reached.
Is snowball sampling qualitative or quantitative?
Both. Qualitative researchers use referral chains to reach people with a shared experience, such as former drug users or migrants, and stop at saturation. Quantitative researchers use them, often as respondent-driven sampling, to estimate prevalence in groups with no sampling frame, which needs coupons, recruitment caps and weighting by network size.
Is a snowball sample the same as a convenience sample?
They overlap. A snowball sample usually starts from seeds chosen for convenience, and it is often classed as a form of convenience sampling. It differs in how it grows: through participants’ own social ties, not the researcher’s access. That adds biases of its own: people refer others like themselves (homophily), protect friends by not referring them (masking), and well-connected people end up over-represented.
What is the difference between purposive and snowball sampling?
In purposive sampling the researcher chooses each participant for characteristics the study needs, while in snowball sampling the first participants (the seeds) recruit the next ones from their own networks, so the researcher controls only the seeds and the criteria. Many qualitative studies combine them: purposively chosen seeds, then referrals to find more people who fit.
Sources
- Biernacki, P., & Waldorf, D. (1981). Snowball sampling: Problems and techniques of chain referral sampling. Sociological Methods & Research, 10(2), 141–163. https://doi.org/10.1177/004912418101000205
- Chung, R. Y.-N., Hui, A. Y.-K., Dong, D., Mak, J. K.-L., & Chia, C. S. M. (2026). Health and social challenges faced by rare disease patients in Hong Kong: A cross-sectional study. Orphanet Journal of Rare Diseases, 21, Article 371. https://doi.org/10.1186/s13023-026-04482-4
- Goodman, L. A. (1961). Snowball sampling. The Annals of Mathematical Statistics, 32(1), 148–170. https://doi.org/10.1214/aoms/1177705148
- Heckathorn, D. D. (1997). Respondent-driven sampling: A new approach to the study of hidden populations. Social Problems, 44(2), 174–199. https://doi.org/10.2307/3096941
- Keygnaert, I., Dialmy, A., Manço, A., Keygnaert, J., Vettenburg, N., Roelens, K., & Temmerman, M. (2014). Sexual violence and sub-Saharan migrants in Morocco: A community-based participatory assessment using respondent driven sampling. Globalization and Health, 10, Article 32. https://doi.org/10.1186/1744-8603-10-32
- Naderifar, M., Goli, H., & Ghaljaie, F. (2017). Snowball sampling: A purposeful method of sampling in qualitative research. Strides in Development of Medical Education, 14(3), e67670. https://doi.org/10.5812/sdme.67670
- Salganik, M. J., & Heckathorn, D. D. (2004). Sampling and estimation in hidden populations using respondent-driven sampling. Sociological Methodology, 34(1), 193–239. https://doi.org/10.1111/j.0081-1750.2004.00152.x
- Wylie, J. L., & Jolly, A. M. (2013). Understanding recruitment: Outcomes associated with alternate methods for seed selection in respondent driven sampling. BMC Medical Research Methodology, 13, Article 93. https://doi.org/10.1186/1471-2288-13-93