[Paper Review] Linked Ego Networks: Improving Estimate Reliability and Validity with Respondent-driven Sampling

Xin Lü|arXiv (Cornell University)|May 9, 2012

HIV, Drug Use, Sexual Risk35 references3 citations

TL;DR

This paper proposes RDSI^{ego}, a novel respondent-driven sampling estimator that integrates ego network data—such as respondents' reports on friends' characteristics—to improve estimation reliability and validity. By incorporating peer recruitment preferences and network structure into the bootstrap procedure, RDSI^{ego} reduces bias to under 2% even under severe violations of RDS assumptions, outperforming traditional RDS estimators that can exhibit biases of 10–20%.

ABSTRACT

Respondent-driven sampling (RDS) is currently widely used for the study of HIV/AIDS-related high risk populations. However, recent studies have shown that traditional RDS methods are likely to generate large variances and may be severely biased since the assumptions behind RDS are seldom fully met in real life. To improve estimation in RDS studies, we propose a new method to generate estimates with ego network data, which is collected by asking RDS respondents about the composition of their personal networks, such as "what proportion of your friends are married?". By simulations on an extracted real-world social network of gay men as well as on artificial networks with varying structural properties, we show that the new estimator, RDSI^{ego} shows superior performance over traditional RDS estimators. Importantly, RDSI^{ego} exhibits strong robustness to the preference of peer recruitment and variations in network structural properties, such as homophily, activity ratio, and community structure. While the biases of traditional RDS estimators can sometimes be as large as 10%~20%, biases of all RDSI^{ego} estimates are well restrained to be less than 2%. The positive results henceforth encourage researchers to collect ego network data for variables of interests by RDS, for both hard-to-access populations and general populations when random sampling is not applicable. The limitation of RDSI^{ego} is evaluated by simulating RDS assuming different level of reporting error.

Motivation & Objective

• To address the high bias and variance in traditional respondent-driven sampling (RDS) estimators due to violated assumptions in real-world settings.
• To improve estimation reliability and validity in hard-to-access populations where sampling frames are unavailable.
• To develop a robust estimator that accounts for differential recruitment and network structural properties such as homophily, activity ratio, and community structure.
• To evaluate the impact of reporting errors on estimator performance and enhance confidence interval coverage through ego network-informed bootstrap methods.
• To encourage the collection of ego network data in RDS studies to improve population estimate accuracy for both hard-to-access and general populations.

Proposed method

• The RDSI^{ego} estimator uses ego network data collected via questions like 'What proportion of your friends are married?' to estimate peer recruitment probabilities.
• It modifies the bootstrap procedure by replacing random recruitment assumptions with ego network-based estimates of recruitment probabilities: $\hat{s}_{AB}^{ego}$ and $\hat{s}_{BA}^{ego}$.
• The method applies a re-weighted random walk (RWRW) framework, using inverse degree weighting to correct for sampling bias.
• Bootstrap resampling is performed using ego network-based recruitment probabilities to generate more accurate confidence intervals.
• Simulations are conducted on real-world (MSM) and artificial networks with varying structural properties to test estimator performance.
• Performance is evaluated using coverage rates of 90% and 95% confidence intervals under random and differential recruitment conditions.

Experimental results

Research questions

• RQ1 How does integrating ego network data into RDS estimation reduce bias under violations of RDS assumptions?
• RQ2 To what extent does RDSI^{ego} improve confidence interval coverage compared to traditional bootstrap methods under differential recruitment?
• RQ3 How robust is RDSI^{ego} to variations in network structure, such as homophily, activity ratio, and community structure?
• RQ4 What is the impact of reporting errors in ego network data on the performance of RDSI^{ego}?
• RQ5 Can RDSI^{ego} maintain low bias and high coverage across diverse network topologies and recruitment patterns?

Key findings

• RDSI^{ego} reduces estimation bias to less than 2% across all simulated conditions, even under severe violations of RDS assumptions.
• Traditional RDS estimators exhibit biases as high as 10–20%, particularly under differential recruitment and non-random network structures.
• The modified bootstrap procedure $BS\textrm{-}ego2$, which uses ego network-based recruitment probabilities, achieves 5–10% higher coverage rates than $BS\textrm{-}ego1$ under differential recruitment.
• On the KOSKK network, $BS\textrm{-}ego2$ increases coverage rates by 8–14% compared to $BS\textrm{-}ego1$ in extreme cases with high differential recruitment.
• Even with reporting errors in ego network data, RDSI^{ego} maintains strong robustness, with bias remaining below 2% in most scenarios.
• Confidence intervals based on $BS\textrm{-}ego2$ show significantly better coverage than $BS\textrm{-}origin$, especially under differential recruitment, where $BS\textrm{-}origin$ coverage drops below 50%.