Essays on causal inference and complex networks

This dissertation is a collection of articles that develop statistical methods for performing causal inference on network data. In bridging these two themes, causal inference and complex networks, the thesis develops four complementary methodological contributions in two main settings that often arise in network data: (i) both the treatment and the outcome are measured at the individual level but the treatment spills over through the network connections; (ii) both the treatment and outcomes are measured at dyadic level. In the first setting, it elaborates innovative techniques for assessing the direct and spillover effects of an intervention in a population of connected units, where the potential outcome of an agent is affected by the treatment status of other interfering agents. In particular, the articles featured in the dissertation expand the existing literature by developing methods that are useful for (i) estimating the effect of an observational multi-valued intervention in a sample of units connected through a weighted network; (ii) detecting and estimating heterogeneous treatment and spillover effects in presence of units who belong to exogenous clusters, and whose interactions are described by cluster-specific networks; (iii) accounting for hidden treatment diffusion processes in a partially unobserved network. In the second setting, the dissertation employs the potential outcomes framework to analyze causal relationships in network formation processes. Specifically, it develops an estimator for the causal effect that the existence of links in a “treatment network” has on the formation of links on an “outcome network,” with both networks being directed.