[2D5-GS-2] Machine learning: Statistical learning

[2D5-GS-2-05] Simultaneous estimation of Conditional Quantile Treatment Effects

〇Kohsuke Uno¹, Koh Takeuchi¹, Hisashi Kashima¹ (1. Kyoto University)

Keywords:Causal Inference, Treatment Effect

The estimation of treatment effect is a method widely used to assess the effect of a specific treatment on a variable in various situations. For example, it is applied to estimate the efficacy of medication on blood glucose levels in diabetic patients. This efficacy is determined by estimating the difference in blood glucose levels between cases of "administration (=treatment)" and "non-administration." Particularly, individual treatment effects for specific units can be precisely measured by estimating the difference at certain quantiles (CQTE) for both cases, allowing for the estimation of the distribution of the outcome variable. However, existing methods often encounter the issue of "crossing quantiles" because they require modeling at each desired quantile level for CQTE estimation. In this study, we propose a method using neural networks for randomized observational data, simultaneously learning all quantiles through a single model for both treatment and control groups, and estimating CQTE at any quantile. This ensures a guaranteed monotonic increase across all quantiles, stabilizing the learning process. Evaluation experiments using artificial data assess the accuracy of distribution estimation for intervention and control groups with different distributions based on the proposed method.