Abstract
Respondent-driven sampling (RDS) is a popular approach to study marginalized or hard-to-reach populations. It collects samples from a networked population by incentivizing participants to refer their friends into the study. One major challenge in analyzing RDS samples is seed bias. Seed bias refers to the fact that when the social network is divided into multiple communities (or blocks), the RDS sample might not provide a balanced representation of the different communities in the population, and such unbalance is correlated with the initial participant (or the seed). In this case, the distributions of estimators are typically non-trivial mixtures, which are determined (1) by the seed and (2) by how the referrals transition from one block to another. This paper shows that (1) block-transition probabilities are easy to estimate with high accuracy, and (2) we can use these estimated block-transition probabilities to estimate the stationary distribution over blocks and thus, an estimate of the block proportions. This stationary distribution on blocks has previously been used in the RDS literature to evaluate whether the sampling process has appeared to 'mix'. We use these estimated block proportions in a simple post-stratified (PS) estimator that greatly diminishes seed bias. By aggregating over the blocks/strata in this way, we prove that the PS estimator is $\sqrt{n}$-consistent under a Markov model, even when other estimators are not. Simulations show that the PS estimator has smaller Root Mean Square Error (RMSE) compared to the state-of-the-art estimators.
Introduction: The RDS Problem
Respondent-Driven Sampling (RDS) is a "snowball" sampling technique designed to reach "hidden" populations—such as drug users or the homeless—where a traditional sampling frame does not exist. The process begins with a few initial members (the "seeds"), who are given coupons to recruit their peers. This process repeats over several "waves," as shown in Figure 1.
While effective for access, RDS suffers from seed bias. If a social network contains "bottlenecks"—insular communities with few inter-group links—the sample can become trapped in the seed's community. Consequently, the final sample may over-represent certain groups, leading to skewed estimates of population characteristics like disease prevalence.
The Post-Stratified (PS) Estimator
To address this, the authors introduce a Post-Stratified (PS) estimator. The core innovation shifts focus from the sample composition itself to the referral patterns.
1. Block Transition Probabilities
Instead of merely counting participants, the PS estimator analyzes the transitions between communities (or "blocks"). It estimates a transition matrix where each entry represents the probability that a member of one group recruits someone from another.
As illustrated in Figure 2, these transition probabilities (e.g., the likelihood of a heroin user recruiting another heroin user vs. someone with a different preference) are remarkably stable. While the overall sample might be biased by the starting seed, these "local" transitions provide a more reliable signal of the underlying network structure.
2. Stationary Distribution and Re-weighting
By treating the transition matrix as a Markov chain, the researchers can calculate its stationary distribution—the equilibrium state the sampling would reach if continued indefinitely. This distribution provides an estimate of the true proportions of each community in the population. The PS estimator then re-weights the sample based on these proportions, combined with a block-wise version of the Volz-Heckathorn (VH) estimator to account for individual popularity (degree).
Key Results and Theoretical Guarantees
The paper provides the first proof that an RDS estimator is $\sqrt{n}$-consistent under a Markov model. This means that as the sample size grows, the error decreases at an optimal rate, even in the presence of severe bottlenecks.
In evaluations using both synthetic networks (generated via Degree-Corrected Stochastic Blockmodels) and real-world data (Add Health school networks), the PS estimator consistently outperformed existing methods:
- Lower Error: It showed smaller Root Mean Square Error (RMSE) and lower absolute bias than the standard VH and state-of-the-art fGLS estimators.
- Robustness: The performance advantage was most pronounced in networks with strong bottlenecks, where other estimators' variances typically explode.
Conclusion
The PS estimator offers a mathematically rigorous solution to seed bias in RDS. While it currently requires known "block labels" (e.g., group membership), future work aims to automate community discovery. By leveraging the stability of local transitions, this method ensures that even if a study starts in the "wrong" place, it can still arrive at the right answer.
Summary updated