Bayesian pairwise estimation under dependent informative sampling (Q1753156)

In informative sampling designs, sample selection from a finite population using the design induces a correlation between study variable and inclusion probabilities of the units. \textit{T. D. Savitsky} and \textit{D. Toth} [Electron. J. Stat. 10, No. 1, 1677--1708 (2016; Zbl 1397.62117)] proposed an automated approach that constructs a sampling-weighted pseudo posterior density by exponentiating each likelihood by a sampling weight inversely proportional to its marginal inclusion probability. Further, Savitsky and Toth's conditions restrict to designs where dependency of pairs of units approaches to zero for large population size \(N\), so as to guarantee consistency of pseudo posterior density estimated from the sample. In this paper, the authors propose an approach where they use a sampling weight that takes into account the pairwise dependency as well. After giving a justification of pairwise weighting, the authors show the consistency of the pairwise pseudo posterior distribution under certain conditions. A simulation study for US National Survey on Drug Use and Health (NSDUH) illustrates their results on calculation of weights. Finally, use of household pairwise weights is shown applicable in surveys such as Job Openings and Labor Turnout Survey (JOLTS).