Primary Submission Category: Interference and Consistency Violations
Generalizing Causal Effects under Partial Interference in Two-Stage Sampling Designs
Authors: Yihan Bao, Laura Forastiere,
Presenting Author: Yihan Bao*
Causal inference under interference is increasingly important in public health and social
science applications, yet most existing methods are either targeted to finite sample
estimands or to super-population estimands implicitly assuming simple random
sampling. In practice, many studies rely on complex multistage survey designs, where
clusters are sampled with unequal probabilities and only a subset of individuals within
each cluster are observed. We develop a general framework under clustered
interference for estimating causal effects and generalizing them from the sample to the
target population when data arise from such two-stage sampling designs. We define
causal estimands as contrasts of average potential outcomes under hypothetical
interventions and derive inverse probability weighted estimators that jointly account for
treatment assignment, cluster-level sampling, and individual-level sampling. We
establish identification conditions under varying assumptions on the interference set,
including settings where potential outcomes depend only on sampled units as well as
more general interference structures involving non-sampled units. When full treatment
information is unavailable, we derive bias expressions and characterize conditions
under which consistency is obtained. Simulation studies illustrate the finite-sample
behavior of the proposed estimators. The methods are applied to estimate the effects of
bed net use on malaria prevalence among children in Uganda.