🙌 New working paper on dealing with overlap in treatment effect estimation. Why it matters in practice: ❌ In clinical studies, some patient groups have low overlap (rare profiles, uncommon treatments) → standard methods struggle to estimate effects reliably ✅ Our 𝐎𝐯𝐞𝐫𝐥𝐚𝐩-𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐑𝐞𝐠𝐮𝐥𝐚𝐫𝐢𝐳𝐚𝐭𝐢𝐨𝐧 (𝐎𝐀𝐑) adjusts for this: stronger regularization where overlap is weak, lighter where overlap is good. This leads to more stable and trustworthy treatment effect estimates, especially in the “hard” regions that clinicians care about most 👉 Our framework works on top of existing CATE meta-learners (parametric or non-parametric), and we also propose debiased variants to preserve orthogonality. In experiments, we show that our OAR yields better performance in low-overlap regions compared to baseline approaches Preprint: https://lnkd.in/dPPUFaQu with Valentyn Melnychuk, Dennis Frauen, Jonas Schweisthal @ Munich Center for Machine Learning
Stefan Feuerriegel thank you, this is another very relevant paper from your group! I will definitely read with interest
Causal Inference & ML | Designing Methods, Building Models & Deploying at Scale for Product Growth
1moI like this. Also, your repository is well organized. Any plans to make this a package?