Robust Causal Inference under Network Uncertainty: Multi-Network Ensemble Approaches

Chao et al. (2023) and Weinstein & Nevo (2023) highlight the bias introduced by mismeasured or misspecified network structures—a pervasive problem in observational social network studies. Rather than relying on a single network (possibly inferred from incomplete or noisy data), this research proposes constructing an ensemble of plausible network representations (e.g., using stochastic block models, ERNMs, chain graphs), then aggregating causal effect estimates across them, perhaps weighting by likelihood or fit. Such an approach could draw on Bayesian model averaging and recent advances in probabilistic bias analysis. The novelty lies in explicitly embracing network uncertainty as a first-class object of inference, rather than a nuisance, offering both more honest uncertainty quantification and improved inference when ground-truth network data is unavailable. This could transform causal analysis in settings—like digital platforms or epidemiology—where the “true” network is always uncertain.

References:

1. Estimation and inference for causal spillover effects in egocentric-network randomized trials in the presence of network membership misclassification.. Ariel Chao, Donna Spiegelman, Ashley L. Buchanan, Laura Forastiere (2023). Biostatistics.
2. Causal inference with misspecified network interference structure. B. Weinstein, D. Nevo (2023).