Abstract
Consider a population of individuals and a network that encodes social connections among them. We are interested in making inference on finite population and super-population estimands that are a function of both individuals' responses and of the network, from a sample. Neither the sampling frame nor the network are available. However, the sampling mechanism implicitly leverages the network to recruit individuals, thus partially revealing social interactions among the individuals in the sample, as well as their responses. This is a common setting that arises, for instance, in epidemiology and healthcare, where samples from hard-to-reach populations are collected using link-tracing mechanisms, including respondent-driven sampling. In this paper, we study statistical properties of popular network sampling mechanisms. We formulate the estimation problem in terms of Rubin's inferential framework to explicitly account for social network structure. We then identify key modeling elements that lead to inferences with good frequentist properties when dealing with data collected through non-ignorable network sampling mechanisms. We demonstrate these methods on a study of the incidence of HIV in Brazil.
Introduction
In public health and sociology, researchers often seek to study "hidden" or hard-to-reach populations, such as individuals at high risk for infectious diseases. Because no complete sampling frame (a list of all population members) exists, traditional random sampling is impossible. Instead, researchers rely on link-tracing mechanisms like Respondent-Driven Sampling (RDS). In RDS, a small set of "seeds" is recruited and given coupons to invite their peers, who in turn invite others, creating a chain-referral process.
While RDS is effective for recruitment, it presents significant statistical challenges. Traditional estimators often assume the sampling process is "ignorable"—meaning that once the number of connections (degree) of an individual is known, the specific path of recruitment does not matter. This paper demonstrates that this assumption is fundamentally flawed and proposes a new Bayesian framework to provide valid inference under non-ignorable conditions.
The Statistical Challenge of Non-Ignorable Sampling
A central contribution of this research is the formal proof (Theorem 2.1) that RDS is a non-ignorable sampling design. In Rubin’s inferential framework, a sampling mechanism is ignorable if the probability of selecting a particular sample does not depend on unobserved data. However, in RDS, the probability of an individual being recruited depends on the unobserved network structure and the recruitment decisions of their peers.
Ignoring this dependency leads to biased estimates and a severe underestimation of uncertainty. The sampling path itself—the sequence of who recruited whom—contains "hidden" information about the network's topology and the distribution of traits within the population. By treating the sampling process as ignorable, standard methods throw away this crucial context.
A Modular Bayesian Framework for Network Inference
To address non-ignorability, we propose a modular Bayesian framework that explicitly models both the population network and the sampling mechanism. The approach is built on several key pillars:
- Bayesian Model Averaging (BMA): Since the unobserved portion of the network is unknown, we treat it as a missing data problem. We use BMA to average over multiple "imputed" versions of the unobserved network, ensuring that our uncertainty about the hidden structure is reflected in the final estimates.
- Markov Random Fields (MRF): Social networks often exhibit homophily, where connected individuals share similar traits. We model individual responses using an MRF to account for this network-induced dependence.
- Non-Ignorable Likelihood: Our likelihood function explicitly incorporates the probability of the observed sampling path, capturing the information inherent in the recruitment process.
Because the normalizing constants (partition functions) in MRFs are computationally difficult to calculate, we employ Annealed Importance Sampling (AIS) within a Metropolis-Hastings MCMC scheme. This ensures that our posterior estimates are both accurate and computationally feasible.
Empirical Evaluation: Simulations and HIV Prevalence
We evaluated the framework using both synthetic and real-world data. In simulations using Erdős-Rényi and Small World network models, the proposed Bayesian estimator consistently outperformed the standard Volz-Heckathorn (VH) estimator. Specifically, our method showed significantly reduced bias and produced credible intervals that were often half the length of traditional bootstrap intervals while maintaining superior coverage.
We applied the method to a real-world study of HIV prevalence among 658 men who have sex with men in Campinas, Brazil. The results demonstrated that our model is robust to variations in assumed population size ($N$), a common sensitivity point in network studies. By accounting for the fact that an individual's HIV status is statistically related to their recruiters' status, the model provided a more nuanced and reliable estimate of community-wide prevalence.
Conclusion and Future Directions
This work establishes that valid inference from network-sampled data requires accounting for the non-ignorable nature of the sampling design. By integrating network modeling with Rubin's inferential framework, we provide a path forward for more accurate public health surveillance in hidden populations. Future research will focus on scaling these computationally intensive methods to larger networks and refining models to account for non-uniform seed selection and differential recruitment activity among different demographic groups.
Summary updated