Primary Submission Category: Machine Learning and Causal Inference
Enhancing Causal Inference: Reducing Uncertainty in Causal Forest Models through Cross-Validation
Authors: Yufan Ji, Abdollah Shafieezadeh, Noah Dormady,
Presenting Author: Yufan Ji*
In data-driven decision-making, understanding causal relationships is essential. Machine learning, especially causal forest models, has transformed causal inference by estimating the Conditional Average Treatment Effect (CATE), improving personalized strategies across sectors like healthcare, education, and energy. Causal forests, an adaptation of random forests, use the potential outcomes framework to assess heterogeneous treatment effects. However, today’s research has primarily focused on accuracy while overlooking the crucial impact of uncertainty on decision making. This is important because large epistemic uncertainties can lead to skewed resource allocation, misinformed prioritization, and suboptimal results in critical areas. This research presents a new method that reduces predictive uncertainty in causal forest models through hyperparameter optimization via Bayesian search and introduces a cross-validation layer to mitigate overfitting. The approach is validated with both synthetic and real data, demonstrating its ability to outperform traditional tools through uncertainty reduction and enhancing decision accuracy. This improvement not only strengthens the reliability of causal estimates but also optimizes decision-making, enabling more effective resource use based on solid causal evidence. These advancements in addressing uncertainty mark a significant contribution to causal inference, with wide-reaching implications for various applications.