New diffusion-based IVQR method beats benchmarks in high-dimensional causality

Researchers Xingdong Feng, Xinhong Jiang, Yuling Jiao, Lican Kang, and Junwei Liu released a paper proposing a new deep learning approach to instrumental variable quantile regression (IVQR) — a method for estimating causal effects when unobserved confounding is present. Their estimator combines two stages: first, a variance-preserving conditional diffusion model learns the joint distribution of outcome and endogenous covariates given an instrument; second, a kernel-smoothed version of the conditional moment condition is minimized over deep neural networks using Monte Carlo sampling. This lets the model recover structural quantile functions without parametric assumptions.

The theory is notably thorough. The authors establish an excess-risk bound for the estimator and derive end-to-end total variation guarantees for the conditional diffusion component under unbounded support. Critically, they account for score estimation error, early stopping, and discretization errors in the diffusion sampling process. Their analysis assumes a polynomial-tail envelope on the data distribution, meaning it applies to heavy-tailed data. As the tail index grows, their bound converges to the minimax-optimal rate of classical nonparametric regression — so the heavy-tailed theory seamlessly reduces to standard light-tailed guarantees.

Empirically, the method outperforms existing nonparametric IVQR techniques in both simulated and real-data tests. The gains become more pronounced as the number of covariates and instruments grows, suggesting that diffusion-based models are especially effective for high-dimensional causal inference. This addresses a major limitation of traditional kernel and spline-based IVQR approaches, which struggle in settings beyond a few dimensions.

For practitioners, this work offers a principled way to estimate quantile treatment effects without relying on strict parametric assumptions. It bridges modern generative modeling with econometrics and could unlock applications in health economics, demand forecasting, and policy evaluation where both confounding and heavy-tailed outcomes are common.