Counterfactual learning in networks : an empirical study of model dependence

Within the potential outcomes framework for causal inference, the choice of unit features and matching algorithms can impact the estimated causal effects, a problem known as model dependence. Here, we look at this problem in the context of observational network data and recently developed network representations within machine learning. By varying node representations, matching models, and methods for causal effect estimation on synthetic and real-world graph datasets, we show experimentally that estimated causal effects can vary significantly, both in sign and magnitude. With this paper, we aim to highlight some of the challenges of estimating causal effects from observational network data and hope to inspire further studies on model dependence in causal inference.