A critical threshold for design effects in network sampling (Q1731771)

The author considers response-driven sampling (RDS) which is a popular technique employed by several national and international organizations to study hard-to-reach populations such as those in HIV research programs. Each participant with an incentive to join the study passes three (to five) referrals, assumed to be a random subset of friends and those that respond will form the first wave of the study. This process iterates until the required sample size is reached or the process discontinues due to non response. Thus RDS can be regarded as a stochastic process on the members of the social network. Though the Markovian model is found to be easily tractable analytically, in practice the estimates of parameters of interest have large variances due to the invalid assumptions. \par The present paper defines a Markov process on a graph \(G\), with node set \(V\) consisting of the set of participants and the edge set \(E\) consisting of the relation of friendship between persons $i$ and $j$ as an edge. This process is indexed by a rooted tree $\mathbb{T}$. For any node $\sigma\in\mathbb{T}$, denote by $\sigma'$ the node one step closer to the root. For each node $i\in G$, let $y(i)$ be a study variable of this node. The problem now is to estimate $\mu=(\sum_{i\in G}y(i))/N$. Under RDS, a closed form expression for $V_1=\mathrm{Var}. (\hat\mu)$ is derived. For an independent random sample $w_1,w_2,\dots, w_n\in G$, if $\hat\mu$ is computed and its variance is denoted by $V_2$, the Design Effect DE of $\hat\mu$ is given by $V_1/V_2$. Next, the asymptotic behaviour of DE for the Galton-Watson process is studied. It is further shown that under certain assumptions on the referral tree, the DE of network sampling has a critical threshold which is a function of the referral rate and the clustering structure in the social network involving $\lambda_2$, the second largest eigenvalue of the Markov transition matrix. It is then shown by resampling of nodes for Markovian model, bootstrap procedures allow computational comparisons with earlier methods in the literature.