Comprehensively evaluating health policies requires extending beyond simple treated versus untreated comparisons. Many policies generate spillover effects that can amplify or offset intended outcomes, while exposure levels vary among individuals within implementing regions. Understanding these variations is crucial for designing effective interventions. Relatedly, identifying the drivers of policy effect heterogeneity is key to uncovering the mechanisms through which policies operate. These analyses are further complicated by several statistical challenges, including confounding from non-randomized policy implementation in real-world settings, violations of standard causal inference assumptions such as interference between units, and spatial and temporal correlations, which standard methods often fail to address.This dissertation develops novel causal inference methodologies to improve policy evaluations using observational data. First, we introduce a doubly robust difference-in-differences (DiD) framework to estimate policy effects in both implementing and indirectly affected neighboring regions while accounting for spatial interference and relaxing assumptions about confounding. Second, we propose multiply robust estimators for causal effect curves of continuous policy exposures, addressing confounding between intervention status, exposure levels, and outcome trends while avoiding strong parametric assumptions. Third, we construct a causal framework for evaluating policy effect heterogeneity by representing relevant research questions as hypothetical interventions and proposing robust estimation approaches, enabling researchers to assess pathways driving effect heterogeneity while addressing confounding and neighborhood dynamics. We evaluate these methods through extensive simulations and apply them to study Philadelphia’s sweetened beverage tax. Our findings highlight the importance of accounting for spillover effects and heterogeneity in policy exposure and effects when designing and assessing public health policies.