Research
Selected publications
-
Biometrics · 2026 · Accepted as a Discussion Paper
Causal Inference with Misspecified Network Interference Structure
Estimation of causal effects under network interference when the assumed interference structure is misspecified. Under a design-based framework, we study the implications of network misspecification on standard estimators and propose novel Network-Misspecification-Robust estimators that remain unbiased if at least one of the observed network is correct. The paper is accepted as a discussion paper ( 1, 2, 3, 4), followed by a rejoinder.
-
arXiv:2505.08395 · 2026 · R&R2, Journal of Machine Learning Research
Bayesian Estimation of Causal Effects Using Proxies of a Latent Interference Network
We introduce a structural causal model framework to estimate causal effects under network interference when the true interference structure is latent and only proxy measurements are available. We develop a Bayesian inference approach that jointly reconstructs the latent network and estimates causal parameters. The posterior distribution is composed of high-dimensional mixed space of discrete and continuous variables, which severly complicates the inference process. We overcome these computational challenges by developing a novel Block Gibbs sampler equipped with Locally Informed Proposals, that efficiently explores the posterior distribution, while ensuring that uncertainty in the network structure is properly propagated to the final causal estimates.
-
Biostatistics · 2026
Sensitivity Analysis for Contamination in Egocentric-Network Randomized Trials with Interference
Egocentric-network randomized trials are frequently used to estimate causal effects under interference. However, contamination between sampled ego-networks if often present and can bias standard direct and indirect effect estimators. To address this issue, we derive bias-corrected estimators and propose a sensitivity analysis framework based on parameters that represent the probability or expected number of missing edges, and can be flexibly specified. This framework is implemented via both grid senstivity analysis and probabilistic bias analysis, providing researchers with a flexible tool to rigorously assess the robustness of their causal estimates to contamination.
-
American Economic Review · 2026
Negative Control Falsification Tests for Instrumental Variable Designs
Instrumental variable identification rests on assumptions that are hard to verify. We develop falsification tests based on negative controls that can detect violations of the assumptions underlying IV designs, providing researchers practical tools to stress-test their assumptions.