Causal machine learning for reliable decision-making

Frauen, Dennis (2026): Causal machine learning for reliable decision-making. Dissertation, LMU München: Faculty of Mathematics, Computer Science and Statistics

Abstract

Data-driven decision-making has become increasingly crucial in various domains, including healthcare, economics, and policy-making. Here, large-scale datasets hold immense potential to inform critical decisions, from optimizing medical treatments to designing effective public policies. For example, in medicine, electronic health records (EHRs) and genomic datasets enable the development of personalized treatment plans by identifying which therapies are most effective for specific patient populations. To fully leverage the potential of large-scale data, it is essential to develop methods that enable reliable decision-making, ensuring that data-driven insights produce robust and actionable outcomes. Unlike predictive tasks, decision making requires understanding the consequences of actions or treatments, which is fundamentally tied to causal inference. However, estimating causal effects from data is challenging for two key reasons: First, methods for estimating causal effects often rely on strong assumptions, such as the absence of unobserved confounding, which may be violated in real-world settings. Second, the fundamental problem of causal inference lies in the fact that we can never observe the outcomes of counterfactual actions, which would have happened if a different decision had been made. Addressing these challenges is crucial to ensure that machine learning methods can be used reliably for decision making. This thesis develops novel methods to address the challenges above, thereby enhancing the reliability of causal machine learning for data-driven decision making. A particular focus is on (i) robustness: we develop methods designed to ensure robustness to potential violations of assumptions, such as unobserved confounding; and (ii) efficiency: we develop methods designed to maximize the information in the available data, ensuring effective causal effect estimation despite the fundamental problem of causal inference. For robustness (i), the first part of the thesis proposes novel machine learning methods for sensitivity analysis and partial identification, specifically designed to mitigate the challenges posed by unobserved confounding. These methods quantify the potential impact of violating standard causal assumptions and provide bounds on the causal effect of interest, offering practitioners tools to make informed decisions even in the face of uncertainty. To address efficiency (ii), the second part of the thesis introduces a family of novel model-agnostic learners, known as meta-learners, for estimating heterogeneous treatment effects. These learners are grounded in statistical efficiency theory and leverage the predictive strength of modern machine learning while maintaining favorable theoretical guarantees. The third and final part of this thesis combines these two dimensions, robustness and efficiency, by developing meta-learners for partial identification. These learners extend the meta-learner framework to settings with unobserved confounding, providing effective estimators for bounds on causal effects that can be used for decision-making when causal assumptions are violated. Together, the contributions of this thesis advance the field of causal machine learning by providing novel methodologies that are both principled and practical, ensuring reliable decision making in complex, real-world settings.

• Item Type: Theses (Dissertation, LMU Munich)
• Keywords: Causal machine learning, treatment effect estimation, policy learning
• Subjects: 000 Computers, Information and General Reference; 000 Computers, Information and General Reference > 004 Data processing computer science
• Faculties: Faculty of Mathematics, Computer Science and Statistics
• Language: English
• Date of oral examination: 24. March 2026
• 1. Referee: Feuerriegel, Stefan
• Signature of the printed copy: 0001/UMC 31904