Respondent-Driven Sampling: Reaching Hidden Populations Through Peer-Referral Chains

Respondent-driven sampling (RDS) is a structured form of chain-referral sampling used to study hidden populations — groups with no public membership list or registry, often because membership is stigmatized, illegal, or otherwise concealed (people who inject drugs, sex workers, undocumented migrants, men who have sex with men in criminalizing jurisdictions, people with a rare and highly stigmatized condition). Because no sampling frame exists, a researcher cannot draw a probability sample the way they would for simple random sampling or stratified sampling. RDS was developed by sociologist Douglas Heckathorn and introduced in “Respondent-Driven Sampling: A New Approach to the Study of Hidden Populations” (Social Problems, 44(2), 174–199, 1997) specifically to close that gap: it recruits through peer referral, like ordinary snowball sampling, but adds a tracked coupon system and a mathematical weighting model that together let researchers calculate population-level estimates with defensible confidence intervals — something an unweighted chain-referral sample cannot do.

This guide covers how the peer-referral chain actually works (seeds, coupons, waves), why RDS requires its own network-size weighting to be more than a convenience sample, the assumptions that weighting depends on, and how to report an RDS study in a Methods section.

RDS vs. Ordinary Snowball Sampling

RDS is best understood as a refinement of snowball sampling, not a wholly separate method. Both start from a small set of initial contacts (seeds) who recruit peers, who recruit further peers, in waves. RDS adds three things ordinary snowball sampling does not have:

• A tracked referral mechanism. Each recruit enters the sample carrying a unique, coded referral coupon that records exactly who recruited them — snowball sampling has no equivalent audit trail.
• A capped, dual incentive structure. Participants are compensated both for completing their own interview and, separately, for each peer they successfully recruit (up to a fixed cap) — this motivates people to actually use their coupons rather than simply declining to refer anyone.
• A mathematical weighting model. Using the recruitment-chain data and each respondent’s reported network size, RDS reweights the sample to correct for uneven recruitment probability — the step covered in detail below, and the reason RDS can support population estimates that plain snowball sampling cannot.

If a study only needs a set of people to interview — not a statistically defensible population estimate — the added cost and complexity of the coupon system and weighting analysis usually is not worth it, and ordinary snowball sampling is the better fit. See Snowball Sampling: Definition, Method, and When to Use It for the base method.

The Mechanics: Seeds, Coupons, and Waves

1. Seed selection

A small number of initial participants (seeds) are purposively recruited — often through an outreach worker, a service agency, or the researcher’s own contacts. Seed diversity matters more in RDS than in plain snowball sampling: because the theoretical justification for RDS weighting depends on the sample composition converging toward equilibrium (see below) regardless of where the chains started, seeds drawn from a single narrow subgroup simply mean more waves are needed before that convergence happens. Seeds are typically excluded from, or flagged separately in, the final weighted analysis, since they were not recruited through the peer-referral process being modeled.

2. The coupon system

Each participant, upon completing their own interview, is given a small, fixed number of uniquely coded referral coupons — commonly three — to pass to other people they personally know who are also members of the hidden population and meet the study’s eligibility criteria. A coupon serves two functions at once: it is the recruitment tool (whoever presents a valid coupon is eligible to enroll), and it is the data record that lets researchers reconstruct the full recruitment tree — who recruited whom, and in what order — which is the raw material the weighting calculation needs.

3. The dual incentive

Participants are compensated for two distinct things: completing their own interview, and each peer they successfully recruit who goes on to complete the study, up to the coupon cap. This secondary, per-recruit incentive is what reliably gets coupons distributed and used — without it, referral chains tend to die out quickly, since an already-compensated participant has little reason to spend effort recruiting.

4. Waves

The sample grows outward in waves (also called recruitment generations): each new recruit becomes, in turn, a potential recruiter for the next wave. Recruitment continues until the target sample size is reached, chains stop producing new eligible contacts, or (checked via the equilibrium diagnostic below) the sample composition has stabilized.

Why RDS Needs Its Own Weighting: The Network-Size Adjustment

Left unweighted, a chain-referral sample — RDS included — is still a convenience sample with a structural bias: people who are more socially connected within the hidden population itself are named as referrals more often, simply because they appear in more of their peers’ personal networks. A person who personally knows thirty other members of the hidden population has far more chances of being handed a coupon than someone who knows two. Left uncorrected, this inflates the apparent share of well-connected people in the sample relative to their true share of the population — the same mechanism that inflates high-degree nodes in any network-based recruitment process.

RDS corrects for this with a single additional survey question: each respondent reports their own personal network size (sometimes called their degree) — the number of other members of the hidden population they personally know. Heckathorn’s follow-up paper, “Respondent-Driven Sampling II: Deriving Valid Estimates from Chain-Referral Samples of Hidden Populations” (Social Problems, 49(1), 11–34, 2002) — the source of what is usually called the RDS-II estimator — formalized how to use that self-reported network size to reweight the sample: each respondent’s contribution to a population estimate is weighted inversely proportional to their reported network size, so someone who reported knowing thirty peers counts for less, per person, than someone who reported knowing two. Because well-connected people were more likely to be sampled in the first place, down-weighting them is what converts the raw referral chain from an uncorrected convenience sample into an estimator whose population-level estimates and standard errors are calculable — provided the assumptions below hold.

Worked Example (Illustrative Simulation — Not a Real Study)

The following is a synthetic, seeded computer simulation built to illustrate the mechanism above, not data from any real study or institution. Every number below was generated by a fixed-seed simulation script (PRNG seed 20260826) and is independently reproducible — it is not an invented or estimated figure.

The simulation defines a hidden population of 500 people split into two strata that differ only in personal network size: 350 people in a lower-connectivity (“isolated”) stratum and 150 in a higher-connectivity (“connected”) stratum — so the true population share of the connected stratum is exactly 30.0%. Five seeds are drawn only from the isolated stratum (a deliberately biased start, to test whether the chain still converges toward the true mix). Each sampled person receives 3 coupons, each used with 55% probability, and new recruits are selected from the remaining population with probability proportional to personal network size — modeling the same over-recruitment-of-the-well-connected mechanism described above. Recruitment ran for 8 waves to reach a target sample of n = 150.

Estimate	Connected-stratum share	Absolute error vs. true value (30.0%)
True population value	30.0%	—
Naive (unweighted) sample share	50.7%	20.7 points
RDS network-size-weighted estimate	22.0%	8.0 points

Two things are worth noting in these results. First, the naive, unweighted sample badly overstates the connected stratum’s true share — exactly the bias the network-size mechanism above predicts, and it happens despite the seeds themselves being drawn entirely from the opposite (isolated) stratum, which shows how quickly degree-based selection can dominate the sample composition once referral chains run for a few waves. Second, the network-size-weighted estimate is substantially closer to the true value but does not recover it exactly — a realistic and expected result, not a flaw in the demonstration: RDS weighting corrects the dominant source of bias, it does not eliminate all sampling variability, and this simulation’s coupon-use probability, degree distribution, and finite sample size all still contribute finite-sample noise around the true value. That gap is exactly why RDS results are reported with confidence intervals rather than as point estimates alone.

The Assumptions RDS’s Weighting Actually Depends On

The network-size weighting above is only defensible if several conditions hold. When they do not, the weighted estimate can still be biased — just less predictably than the raw, unweighted sample:

• Reciprocal ties. If person A considers person B a network member, B must also consider A one. One-directional relationships (a celebrity known to many who does not know them back) break the degree-based logic.
• Random recruitment within a personal network. A participant is assumed to pick who to refer roughly at random among the peers they could refer, not systematically favor a particular subgroup (e.g., always referring people who share their own specific characteristics). Real behavior violates this to some degree — it is a source of homophily bias, discussed below.
• Sampling with replacement (approximately). The mathematical model treats recruitment as if drawn with replacement from the network. This is an approximation, reasonable when the sample is a small fraction of a large population, and increasingly strained as the sample approaches the true population size.
• A single, connected network. If the hidden population actually consists of separate, non-overlapping social clusters, chains started in one cluster may never reach the others — no amount of weighting recovers a subgroup the referral chains never touched.
• Accurate self-reported network size. Respondents must be able to estimate their own personal network size within the hidden population reasonably accurately. Round-number heaping (reporting 10, 20, or 50 rather than a precise count) and genuine uncertainty about who else belongs to a stigmatized population are both real, documented threats to this assumption.

Equilibrium is the standard diagnostic for whether the chain has run long enough for these assumptions to plausibly be doing their job: researchers plot the cumulative sample composition (e.g., cumulative share of a given characteristic) against wave number and check whether it has stabilized, independent of the seeds’ own characteristics. In the worked example above, the connected-stratum share moved from 23.1% after wave 1 to 50.7% by wave 8 — still shifting, not yet flat — which is itself informative: a real RDS study seeing that pattern would recruit further waves before treating the estimate as stable, rather than stopping early.

When RDS Is (and Isn’t) the Right Choice

RDS is a strong fit when all three of the following hold: the population is genuinely hidden (no usable sampling frame), members plausibly know and can refer other members (a real social network exists among them), and the study needs a population-level estimate with a computable confidence interval — not just a set of interview subjects. This is why RDS is the standard method in public-health surveillance of populations such as people who inject drugs or men who have sex with men, where behavioral prevalence estimates directly inform resource allocation.

RDS is usually the wrong choice when: the population, while hard to reach, is not meaningfully networked (members mostly don’t know each other, so referral chains die immediately) — purposive sampling or convenience sampling may fit better; the study is exploratory or qualitative and does not need a weighted population estimate, in which case plain snowball sampling gets the same interview access at lower cost and complexity; or a real sampling frame actually exists, in which case a probability method (simple random sampling, stratified sampling, cluster sampling) will always be preferable to any referral-based design.

Reporting RDS in a Methods Section

A methods section describing an RDS study should disclose, at minimum:

• Number of seeds, how they were selected, and their observable characteristics (to allow readers to assess seed diversity).
• The coupon cap per participant and the incentive structure, including the value of the primary and secondary (per-recruit) incentives.
• Number of recruitment waves and the final sample size.
• The equilibrium diagnostic used and whether the sample reached it before analysis.
• The specific weighting estimator and software used (for example, RDS Analyst or the RDS package in R) and the exact network-size question used to derive weights.
• An explicit statement that findings are not statistically generalizable beyond the study population unless the assumptions above are defensible for that population — RDS narrows, but does not eliminate, the interpretive caveats inherent to any non-probability design.

See Sampling Bias: Types, Causes, and How to Detect It for the broader vocabulary researchers use to describe and disclose these limitations, and Sampling Methods: Probability and Non-Probability Types Explained for how RDS fits into the full landscape of sampling designs.

Frequently Asked Questions

What is respondent-driven sampling?

Respondent-driven sampling (RDS) is a chain-referral sampling method for studying hidden populations that have no public sampling frame. Existing participants recruit peers using a tracked coupon system, and the resulting sample is reweighted using each participant’s reported personal network size, allowing researchers to calculate population-level estimates with confidence intervals.

How is RDS different from snowball sampling?

Snowball sampling is the general chain-referral method: participants recruit peers, with no formal tracking or weighting. RDS adds a tracked coupon system, a capped dual-incentive structure, and a mathematical weighting model based on network size — the combination that lets RDS support statistically defensible population estimates, which ordinary snowball sampling cannot.

How many coupons does each RDS participant get?

There is no fixed universal number, but three is the most commonly used cap in practice — enough to sustain chain growth without letting any single participant dominate the recruitment tree. The exact cap is a study design choice, balanced against budget and the desired number of waves.

What is the network-size adjustment in RDS?

It is the weighting step that corrects for the fact that people with larger personal networks (within the hidden population) are more likely to be recruited through chain referral. Each respondent reports their own network size, and their contribution to the final population estimate is weighted inversely to that number, down-weighting well-connected respondents who were structurally over-sampled.

Can RDS results be generalized to the full hidden population?

Only under specific, checkable assumptions: reciprocal network ties, roughly random recruitment within each participant’s personal network, a single connected network (not fragmented sub-populations chains never reach), and accurate self-reported network size. When those hold, RDS supports population-level estimates with confidence intervals; when they don’t, the weighted estimate can still be biased, just less predictably than an unweighted sample.

What software is used to analyze RDS data?

The two most commonly used tools are RDS Analyst (a free, purpose-built desktop application) and the RDS package in R, both of which implement network-size weighting estimators and compute the associated confidence intervals from the recruitment-chain data.