A1484
Title: Bayesian estimation of causal effects from multiple datasets using structural causal models
Authors: Shunsuke Horii - Waseda University (Japan) [presenting]
Yoichi Chikahara - NTT Communication Science Laboratories (Japan)
Abstract: A novel Bayesian approach is proposed for estimating causal effects from experimental data with multiple interventions. The method builds on structural causal models (SCMs) by calculating the posterior distributions of the parameters of structural equations derived from multiple datasets. Using these posterior distributions, the posterior distributions of the causal estimands are computed. Unlike traditional approaches that estimate only the average causal effect (ACE) from multiple data sources, this method offers two key advantages. First, it achieves Bayesian optimality, leading to improved estimation performance across a range of scenarios. Second, the approach extends beyond the estimation of ACE to allow for the estimation of modified causal effects, such as those conditioned on specific covariates or interventions. This added flexibility enhances the interpretability and applicability of causal estimates in real-world settings where effect modification is crucial. Through experiments on simulated data, it is demonstrated that the Bayesian framework outperforms existing methods in terms of estimation accuracy and provides more nuanced insights into causal relationships.
