Abstract
Researchers have extensively explored networked epidemiology, and with the recent outbreaks of
COVID-19, there has been a growing interest in understanding not only correlations but also the
causal structure underlying disease transmission. In this thesis, the term “causal” refers to two
distinct notions, depending on the research question. For research questions 1 and 2, “causal effect”
denotes interventional causality in the sense of the Neyman-Rubin potential outcomes framework and
structural causal models. For research question 3, “causal influence” refers to predictive, Granger
style causality, which captures whether the past of one time series improves the prediction of another.
We make this distinction explicit throughout the thesis.
Spatiotemporal networks, such as those arising in epidemiology and mobility systems, pose sig
nificant challenges for both causal inference and forecasting. In these dynamic settings, a unit’s
outcome may depend on its own treatment as well as the treatments of neighboring units, lead
ing to violations of classical assumptions such as unconfoundedness and no interference. Moreover,
treatments are often entangled across connected units, and interventions in one region may generate
indirect, spillover effects in others.
This thesis addresses three complementary methodological challenges in this domain: (i) esti
mating interventional causal effects under entangled treatment assignments and unobserved con
founding using a graph-based instrumental variable framework; (ii) quantifying spillover effects in
epidemic networks via a directional graph neural network (Dir-GNN) estimator developed under an
interventional causal framework; and (iii) constructing real-time forecasting models that leverage
time-varying spatiotemporal graphs encoding Granger-style predictive causal relationships derived
from multivariate transfer entropy.
Because these questions rely on different causal formalisms, this dissertation does not seek to
unify interventional and predictive causality. Instead, it clarifies their distinct roles in epidemic analysis:
interventional causality enables counterfactual reasoning about policy effects, while predictive
causality reveals directional dependencies useful for forecasting. Taken together, the contributions
provide a comprehensive methodological toolkit for analyzing and forecasting epidemic processes on
dynamic spatial networks under the appropriate causal interpretation.
