Estimating Causal Effects in Observational Data via Double ML Methods

Thesis Abstract

Estimating causal effects from observational data remains a central challenge in empirical research, where treatment assignment is not randomized and confounding factors may bias conventional estimators. This study addresses the problem by applying double machine learning (DML) methods to consistently estimate causal effects in the presence of high-dimensional covariates, aiming to improve policy-relevant inferences in non-experimental settings. The objective is to quantify the average treatment effect (ATE) and conditional average treatment effects (CATE) of a binary intervention on a real-world outcome, while rigorously controlling for a rich set of confounders through machine learning-based nuisance parameter estimation. The research advances by integrating theoretical developments in causal inference with robust empirical validation across multiple contexts, thereby delineating the conditions under which DML yields unbiased and efficient estimates in observational data. A quasi-experimental research design is employed, drawing on a temporally bounded observational dataset from a large public health program implemented across 12 metropolitan regions over a four-year period. The population comprises adult residents eligible for the program, with a target sample size of 8,000 individuals who encountered the treatment and 8,000 comparable controls identified via propensity score–based matching on pre-treatment covariates. Data collection utilizes administrative records linked to survey instruments, including detailed demographic information, prior health outcomes, socioeconomic indicators, geographic mobility, and program participation histories. The study also assembles a comprehensive covariate set (n ? 150) drawn from health records, social services data, and regional economic indicators to satisfy the overlap/positivity assumptions required for DML estimation. Data quality is ensured through validation checks, harmonization procedures, and imputation of missing values where appropriate, guided by established best practices in causal machine learning. The analysis proceeds in two stages. First, nuisance functions for the propensity score and the outcome regression are estimated using a diverse library of machine learning algorithms, including gradient boosting (XGBoost), neural networks, Lasso and ridge regression, and random forests, with cross-validated hyperparameters. Second, the DML estimator is applied to obtain unbiased estimates of the ATE and CATE, leveraging Neyman-orthogonal moments to mitigate regularization bias. Complementary analyses include traditional regression and propensity score methods for robustness checks, as well as sensitivity analyses to assess the impact of potential unmeasured confounding using Rosenbaum bounds. Model diagnostics focus on the overlap condition, covariate balance after weighting, and the stability of estimates across alternative ML specifications. The analysis is conducted in R and Python, with reproducible scripts and a registered protocol to ensure transparency and replicability. Key findings are anticipated to indicate a statistically significant positive effect of the program on the primary health outcome, with estimated ATEs exhibiting confidence intervals that remain robust under alternative ML specifications and sample splits. Heterogeneous treatment effects are expected to reveal that the impact is larger among individuals with lower baseline health metrics and higher exposure to social determinants of disadvantage, demonstrating the value of CATE estimation for targeted policy design. The study also expects that the DML estimates will exhibit tighter standard errors and reduced bias compared with conventional regression-based approaches, especially in the presence of high-dimensional controls and complex nonlinear relationships among covariates. This research contributes to knowledge by operationalizing double machine learning in a large-scale public policy context, illustrating practical procedures for high-dimensional nuisance estimation, and providing empirical evidence on the reliability of DML under real-world data constraints. The integration of causal inference theory with machine learning in a field setting advances methodological rigor for observational studies and informs policymakers about which subpopulations derive the greatest benefit from intervention programs. The main conclusions are that, when applied with careful adherence to overlap conditions and rigorous cross-validation, DML yields credible, policy-relevant causal estimates in observational data. Recommendations include adopting DML as a standard analytic approach in evaluation studies, expanding data linkage to enhance covariate richness, and conducting pre-registered analytic plans to strengthen inference in non-experimental settings. Future work could extend the framework to multi-valued treatments and longitudinal outcomes to capture dynamic treatment effects over time.