Accuracy of Respondent Driven Sampling estimates: A test using the 2021 English census data as a benchmark

Background: Respondent Driven Sampling (RDS) is a widely used chain‑referral method for surveying hidden or hard‑to‑reach populations. While it has theoretical advantages over other non‑probabilistic methods, there is limited empirical evidence on the accuracy of its estimators when benchmarked against complete population data.Methods: We conducted a GPS‑verified, smartphone‑based RDS survey in York, England, in September–October 2024. The target sample size was 800 residents aged 16 or over; 807 eligible responses were obtained from six recruitment seeds over up to 44 waves. The survey collected demographic, socio‑economic, and health information aligned with the 2021 Census, along with personal network size and some network composition data. RDS‑I, RDS‑II, and Successive Sampling estimators were applied, and results were compared with Census 2021 benchmarks.Results: For many variables, RDS estimates diverged markedly from census values, and 95% confidence intervals often failed to include the true value. Deviations from the at‑random recruitment assumption were evident, particularly for sex and ethnicity, and appeared to be a major source of bias. Incorporating self‑reported network composition reduced bias in the estimate of the variable for which network composition information was collected.Conclusions: In this real‑world test, standard RDS estimators did not produce consistently accurate estimates. Violations of the random recruitment assumption were a likely cause. Network composition‑based estimation offers a promising alternative but requires further testing and methodological refinement. Enhancing RDS systems to monitor subgroup representation in real time and applying weighting adjustments to address complex recruitment homophily could improve estimator performance.