Primary Submission Category: Causal Inference and Bias/Discrimination
Nonparametric Inference on Dose-Response Curves Without Positivity Condition
Authors: Yikun Zhang, Yen-Chi Chen, Alexander Giessing,
Presenting Author: Yikun Zhang*
This paper presents a novel integral estimator for estimating the dose-response curve in the presence of spatial confounding without requiring the commonly assumed positivity condition. Our approach involves estimating the derivative of the treatment effect and integrating it to address the bias resulting from the lack of positivity condition. We also provide a fast and reliable numerical recipe for approximating our estimator and derive related asymptotic theory.
To account for uncertainty, we propose a bootstrap method that generates a simultaneous confidence band for the dose-response curve. Additionally, we propose an inverse probability weighting approach for estimating the derivative effect under the violation of positivity conditions, which exhibits connections to the support estimation problem.