Statistical models can improve causal inference without requiring correct specification. This dissertation develops and applies model-assisted methods for causal estimation using experimental and observational data in political science. Each chapter addresses a different aspect of causal inference, demonstrating how auxiliary predictions can be incorporated into standard frameworks to gain efficiency, reduce bias, or diagnose existing procedures. In Chapter 1, I introduce an adaptive stratification procedure for randomized experiments. While researchers often use blocking to increase efficiency, conventional practices fail to exploit the predictive relationship between covariates and potential outcomes. I propose an adaptive procedure for paired designs that allows observations to be rematched across different batches and introduce a stratified estimator that allows for nonparametric covariate adjustment. I show that stratification complements rather than substitutes for regression adjustment, insuring against adjustment error. Simulations using synthetic and experimental data demonstrate that the gains in precision can be substantial. In Chapter 2, I apply model-assisted estimation to a two-dimensional regression discontinuity design (2D-RDD) to study whether green tax credits targeted at economically vulnerable areas can shift voter support. Focusing on the Energy Community Tax Credit Bonus (ECTCB) under the 2022 Inflation Reduction Act, I estimate its impact on the Democratic share of the two-party vote in the 2024 presidential election. The analysis provides causal evidence that the policy failed to generate positive electoral feedback; rather than increasing support for the incumbent, the results indicate a small negative effect (point estimate: -0.39%; 95% confidence interval: [-0.78%, -0.02%]). Methodologically, this chapter develops an improved 2D-RDD estimator that integrates bagging with bootstrap-based inference. In Chapter 3, I examine the properties of covariate adjustment in randomized experiments, specifically the popular Lin estimator. I show that it can produce confidence intervals with below-nominal coverage. For superpopulation inference, the Lin estimator with HC2 standard errors is anticonservative because it omits the sampling variance component. In finite-population settings, HC2 standard errors become inconsistent when the number of covariates is large relative to the size of the smaller treatment arm. I characterize the conditions under which common adjustment methods remain valid and provide a decision flowchart. To address these issues, I propose using cross-estimation with data-adaptive methods and a score-based variance estimator.
