ESSAYS ON SPATIAL AND NETWORK CAUSAL INFERENCE FOR SOCIAL SCIENCES

Abstract

Over the last few decades, causal inference has transformed the quantitative social sciences. However, this revolution still has a long way to go; most existing methodologies assume away the spatial and network interaction of people, an essential element of the social sciences. In this dissertation, I tackle this long-standing methodological challenge to improve the spatial and network causal inference for the social sciences.

The first essay considers the identification of causal diffusion effects, also known as peer and contagion effects. Despite its significant importance in the social sciences, causal diffusion analysis has been challenging due to contextual confounding and homophily bias. Using a general class of dynamic causal directed acyclic graphs, I introduce a new assumption of structural stationarity, under which I develop a placebo test to detect a wide range of biases, including the two types mentioned above. I then propose a difference-in-differences style estimator that can correct biases under an additional assumption.

The second essay proposes a causal inference framework for policy diffusion. Political scientists have studied how policies diffuse across states and countries. Even though the issue of causal inference has received limited attention in the policy diffusion literature, this type of spatial and network analysis is precisely the area where omitted variable bias is common and large. I extend the placebo test and the bias-corrected estimator developed in the first essay to survival analysis. I also provide new sensitivity analysis methods for assessing the robustness of conclusions to potential violations of a key assumption.

The third essay examines the identification of spillover effects. When people can interact with each other, the outcome of one unit may be affected by others’ treatments. In many social science experiments, such spillover effects may occur through multiple networks, and yet, it is often difficult to measure all relevant networks. I show that, unlike conventional omitted variable bias, bias due to unobserved networks remains even in randomized experiments. I develop parametric and nonparametric sensitivity analysis methods to assess the potential influence of unobserved networks on causal findings.