Uniform inference for kernel instrumental variable regression

Instrumental variable regression is a foundational tool for causal analysis across the social and biomedical sciences. Recent advances use kernel methods to estimate nonparametric causal relationships, with general data types, while retaining a simple closed-form expression. Empirical researchers ultimately need reliable inference on causal estimates; however, uniform confidence sets for the method remain unavailable. To fill this gap, we develop valid and sharp confidence sets for kernel instrumental variable regression, allowing general nonlinearities and data types. Computationally, our bootstrap procedure requires only a single run of the kernel instrumental variable regression estimator. Theoretically, it relies on the same key assumptions. Overall, we provide a practical procedure for inference that substantially increases the value of kernel methods for causal analysis.

💡 Research Summary

This paper addresses a critical gap in the literature on kernel instrumental variable (KIV) regression: the lack of uniform confidence bands for the non‑parametric estimator. While recent work has shown that KIV enjoys closed‑form solutions and can handle complex data types (graphs, sequences, preference rankings), practitioners have been unable to quantify uncertainty in a way that is both computationally feasible and theoretically sound. The authors propose a bootstrap‑based inference procedure that requires only a single run of the KIV estimator and delivers confidence sets that are both valid (coverage at least the nominal level) and sharp (not overly conservative).

The methodological contribution rests on three standard NPIV assumptions—low effective dimension (spectral decay of the covariance operators), source condition (smoothness of the true structural function in the RKHS), and link condition (smoothness of the conditional expectation of the structural function given the instrument). Under these conditions, the authors decompose the estimation error into a residual term, a pre‑Gaussian term, and a bias term. The residual term is shown to be negligible at the √n rate, and the bias term vanishes thanks to the source and link conditions. The central challenge is the pre‑Gaussian term, which they express as a sum of mean‑zero random functions U_i.

To approximate the distribution of this term, the paper introduces anti‑symmetric Gaussian multipliers, a novel twist on the classic multiplier bootstrap that cancels the complex bias inherent in ill‑posed inverse problems. By analyzing the local width σ²(T,m) of the modified covariance operator T = S* S⁻¹_z S, they establish non‑asymptotic Gaussian couplings (building on Zaitsev (1987) and Buzun et al. (2022)) and corresponding bootstrap couplings (extending Freedman (1981) and Chernozhukov et al. (2014, 2016)). The key condition is that σ²(T,m) → 0 as m → ∞, which is guaranteed by the low‑effective‑dimension assumption.

The bootstrap algorithm proceeds as follows: (1) compute the KIV estimator once using the kernel matrices K_{XX} and K_{ZZ}; (2) draw i.i.d. standard normal vectors and form anti‑symmetric multipliers; (3) construct the bootstrap statistic B_n = n^{1/2} Σ_i g_i U_i, where g_i are the multipliers and U_i are the influence functions defined in the paper; (4) obtain critical values from the empirical distribution of sup_x |B_n(x)|. Importantly, no additional kernel evaluations or matrix inversions are required beyond the initial KIV fit, making the procedure computationally cheap.

Theoretical results prove that the resulting confidence set \hat C_n satisfies τ‑validity with τ = O(n^{-1}) and (δ,τ)‑sharpness with δ = O(1/ log n). Thus the coverage error shrinks at a near‑parametric rate while the band width is not excessively inflated. The authors also discuss how their techniques extend to other ill‑posed inverse problems, offering a blueprint for uniform inference beyond the specific KIV setting.

The paper situates its contribution relative to prior work on kernel ridge regression (where the inverse problem is absent) and series‑based NPIV estimators (which are limited to low‑dimensional Euclidean covariates). By leveraging the flexibility of kernels, the proposed method can be applied to non‑standard data such as preference rankings, where the feature space may be factorial in size but the kernel captures similarity efficiently.

In summary, this work delivers a practically implementable, theoretically rigorous uniform inference method for kernel instrumental variable regression. It bridges the gap between the expressive power of kernel‑based non‑parametric causal estimation and the need for reliable uncertainty quantification, thereby expanding the toolkit available to researchers in economics, epidemiology, and other fields that rely on causal inference from observational data.